Merge pull request #267 from pclove1:ocl_hough

modules/ocl/doc/image_processing.rst

@@ -329,3 +329,38 @@ Interpolate frames (images) using provided optical flow (displacement field).
     :param newFrame: Output image.
 
     :param buf: Temporary buffer, will have width x 6*height size, CV_32FC1 type and contain 6 oclMat: occlusion masks for first frame, occlusion masks for second, interpolated forward horizontal flow, interpolated forward vertical flow, interpolated backward horizontal flow, interpolated backward vertical flow.
+
+
+ocl::HoughCircles
+-----------------
+Finds circles in a grayscale image using the Hough transform.
+
+.. ocv:function:: void ocl::HoughCircles(const oclMat& src, oclMat& circles, int method, float dp, float minDist, int cannyThreshold, int votesThreshold, int minRadius, int maxRadius, int maxCircles = 4096)
+
+.. ocv:function:: void ocl::HoughCircles(const oclMat& src, oclMat& circles, HoughCirclesBuf& buf, int method, float dp, float minDist, int cannyThreshold, int votesThreshold, int minRadius, int maxRadius, int maxCircles = 4096)
+
+    :param src: 8-bit, single-channel grayscale input image.
+
+    :param circles: Output vector of found circles. Each vector is encoded as a 3-element floating-point vector  :math:`(x, y, radius)` .
+
+    :param method: Detection method to use. Currently, the only implemented method is  ``CV_HOUGH_GRADIENT`` , which is basically  *21HT* , described in  [Yuen90]_.
+
+    :param dp: Inverse ratio of the accumulator resolution to the image resolution. For example, if  ``dp=1`` , the accumulator has the same resolution as the input image. If  ``dp=2`` , the accumulator has half as big width and height.
+
+    :param minDist: Minimum distance between the centers of the detected circles. If the parameter is too small, multiple neighbor circles may be falsely detected in addition to a true one. If it is too large, some circles may be missed.
+
+    :param cannyThreshold: The higher threshold of the two passed to  the :ocv:func:`ocl::Canny`  edge detector (the lower one is twice smaller).
+
+    :param votesThreshold: The accumulator threshold for the circle centers at the detection stage. The smaller it is, the more false circles may be detected.
+
+    :param minRadius: Minimum circle radius.
+
+    :param maxRadius: Maximum circle radius.
+
+    :param maxCircles: Maximum number of output circles.
+
+    :param buf: Optional buffer to avoid extra memory allocations (for many calls with the same sizes).
+
+.. note:: Currently only non-ROI oclMat is supported for src.
+.. seealso:: :ocv:func:`HoughCircles`
+

modules/ocl/include/opencv2/ocl/ocl.hpp

@@ -41,8 +41,8 @@
 //
 //M*/
 
-#ifndef __OPENCV_GPU_HPP__
-#define __OPENCV_GPU_HPP__
+#ifndef __OPENCV_OCL_HPP__
+#define __OPENCV_OCL_HPP__
 
 #include <memory>
 #include <vector>
@@ -827,6 +827,22 @@ namespace cv
 
         };
 
+        ///////////////////////////////////////// Hough Transform /////////////////////////////////////////
+        //! HoughCircles
+        struct HoughCirclesBuf
+        {
+            oclMat edges;
+            oclMat accum;
+            oclMat srcPoints;
+            oclMat centers;
+            CannyBuf cannyBuf;
+        };
+
+        CV_EXPORTS void HoughCircles(const oclMat& src, oclMat& circles, int method, float dp, float minDist, int cannyThreshold, int votesThreshold, int minRadius, int maxRadius, int maxCircles = 4096);
+        CV_EXPORTS void HoughCircles(const oclMat& src, oclMat& circles, HoughCirclesBuf& buf, int method, float dp, float minDist, int cannyThreshold, int votesThreshold, int minRadius, int maxRadius, int maxCircles = 4096);
+        CV_EXPORTS void HoughCirclesDownload(const oclMat& d_circles, OutputArray h_circles);
+
+    
         ///////////////////////////////////////// clAmdFft related /////////////////////////////////////////
         //! Performs a forward or inverse discrete Fourier transform (1D or 2D) of floating point matrix.
         //! Param dft_size is the size of DFT transform.
@@ -1746,4 +1762,4 @@ namespace cv
 #if _MSC_VER >= 1200
 #pragma warning( pop)
 #endif
-#endif /* __OPENCV_GPU_HPP__ */
+#endif /* __OPENCV_OCL_HPP__ */

modules/ocl/perf/perf_hough.cpp

@@ -0,0 +1,98 @@
+/*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, Multicoreware, Inc., all rights reserved.
+// Copyright (C) 2010-2012, 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 oclMaterials 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*/
+
+#include "precomp.hpp"
+
+#ifdef HAVE_OPENCL
+
+using namespace cv;
+using namespace perf;
+
+//////////////////////////////////////////////////////////////////////
+// HoughCircles
+
+typedef std::tr1::tuple<cv::Size, float, float> Size_Dp_MinDist_t;
+typedef perf::TestBaseWithParam<Size_Dp_MinDist_t> Size_Dp_MinDist;
+
+PERF_TEST_P(Size_Dp_MinDist, OCL_HoughCircles,
+            testing::Combine(
+                testing::Values(perf::sz720p, perf::szSXGA, perf::sz1080p),
+                testing::Values(1.0f, 2.0f, 4.0f),
+                testing::Values(1.0f, 10.0f)))
+{
+    const cv::Size size = std::tr1::get<0>(GetParam());
+    const float dp      = std::tr1::get<1>(GetParam());
+    const float minDist = std::tr1::get<2>(GetParam());
+
+    const int minRadius = 10;
+    const int maxRadius = 30;
+    const int cannyThreshold = 100;
+    const int votesThreshold = 15;
+
+    cv::RNG rng(123456789);
+
+    cv::Mat src(size, CV_8UC1, cv::Scalar::all(0));
+
+    const int numCircles = rng.uniform(50, 100);
+    for (int i = 0; i < numCircles; ++i)
+    {
+        cv::Point center(rng.uniform(0, src.cols), rng.uniform(0, src.rows));
+        const int radius = rng.uniform(minRadius, maxRadius + 1);
+
+        cv::circle(src, center, radius, cv::Scalar::all(255), -1);
+    }
+    
+    cv::ocl::oclMat ocl_src(src);
+    cv::ocl::oclMat ocl_circles;
+
+    declare.time(10.0).iterations(25);
+    
+    TEST_CYCLE()
+    {
+        cv::ocl::HoughCircles(ocl_src, ocl_circles, CV_HOUGH_GRADIENT, dp, minDist, cannyThreshold, votesThreshold, minRadius, maxRadius);
+    }
+    
+    cv::Mat circles(ocl_circles);    
+    SANITY_CHECK(circles);
+}
+
+#endif // HAVE_OPENCL

modules/ocl/src/hough.cpp

@@ -0,0 +1,416 @@
+/*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) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage 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*/
+
+#include "precomp.hpp"
+
+using namespace std;
+using namespace cv;
+using namespace cv::ocl;
+
+#if !defined (HAVE_OPENCL) 
+
+void cv::ocl::HoughCircles(const oclMat&, oclMat&, int, float, float, int, int, int, int, int) { throw_nogpu(); }
+void cv::ocl::HoughCircles(const oclMat&, oclMat&, HoughCirclesBuf&, int, float, float, int, int, int, int, int) { throw_nogpu(); }
+void cv::ocl::HoughCirclesDownload(const oclMat&, OutputArray) { throw_nogpu(); }
+
+#else /* !defined (HAVE_OPENCL) */
+
+#define MUL_UP(a, b) ((a)/(b)+1)*(b)
+
+namespace cv { namespace ocl {
+    ///////////////////////////OpenCL kernel strings///////////////////////////
+    extern const char *imgproc_hough;
+
+    namespace hough
+    {
+        int buildPointList_gpu(const oclMat& src, oclMat& list);
+        void circlesAccumCenters_gpu(const unsigned int* list, int count, const oclMat& dx, const oclMat& dy, oclMat& accum, int minRadius, int maxRadius, float idp);
+        int buildCentersList_gpu(const oclMat& accum, oclMat& centers, int threshold);
+
+        int circlesAccumRadius_gpu(const oclMat& centers, int centersCount,
+                                   const oclMat& list, int count,
+                                   oclMat& circles, int maxCircles,
+                                   float dp, int minRadius, int maxRadius, int threshold);
+    }
+}}
+
+
+
+//////////////////////////////////////////////////////////
+// common functions
+
+namespace cv { namespace ocl { namespace hough
+{
+    int buildPointList_gpu(const oclMat& src, oclMat& list)
+    {
+        const int PIXELS_PER_THREAD = 16;
+
+        int totalCount = 0;
+        int err = CL_SUCCESS;
+        cl_mem counter = clCreateBuffer(src.clCxt->impl->clContext,
+                                        CL_MEM_COPY_HOST_PTR,  
+                                        sizeof(int),
+                                        &totalCount,   
+                                        &err);
+        openCLSafeCall(err);
+
+        const size_t blkSizeX = 32;
+        const size_t blkSizeY = 4;
+        size_t localThreads[3] = { blkSizeX, blkSizeY, 1 };
+
+        const int PIXELS_PER_BLOCK = blkSizeX * PIXELS_PER_THREAD;
+        const size_t glbSizeX = src.cols % (PIXELS_PER_BLOCK) == 0 ? src.cols : MUL_UP(src.cols, PIXELS_PER_BLOCK);
+        const size_t glbSizeY = src.rows % blkSizeY == 0 ? src.rows : MUL_UP(src.rows, blkSizeY);      
+        size_t globalThreads[3] = { glbSizeX, glbSizeY, 1 };
+
+        vector<pair<size_t , const void *> > args;
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&src.data ));
+        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 *)&src.step ));
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&list.data ));
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&counter ));
+
+        openCLExecuteKernel(src.clCxt, &imgproc_hough, "buildPointList", globalThreads, localThreads, args, -1, -1);
+        openCLSafeCall(clEnqueueReadBuffer(src.clCxt->impl->clCmdQueue, counter, CL_TRUE, 0, sizeof(int), &totalCount, 0, NULL, NULL));  
+        openCLSafeCall(clReleaseMemObject(counter));
+        
+        return totalCount;
+    }    
+}}}
+
+//////////////////////////////////////////////////////////
+// HoughCircles
+
+namespace cv { namespace ocl { namespace hough
+{
+    void circlesAccumCenters_gpu(const oclMat& list, int count, const oclMat& dx, const oclMat& dy, oclMat& accum, int minRadius, int maxRadius, float idp)
+    {
+        const size_t blkSizeX = 256;
+        size_t localThreads[3] = { 256, 1, 1 };
+        
+        const size_t glbSizeX = count % blkSizeX == 0 ? count : MUL_UP(count, blkSizeX);
+        size_t globalThreads[3] = { glbSizeX, 1, 1 };
+
+        const int width  = accum.cols - 2;
+        const int height = accum.rows - 2;
+
+        vector<pair<size_t , const void *> > args;
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&list.data ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&count ));
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&dx.data ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&dx.step ));
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&dy.data ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&dy.step ));        
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&accum.data ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&accum.step ));        
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&width ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&height ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&minRadius));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&maxRadius));
+        args.push_back( make_pair( sizeof(cl_float), (void *)&idp));
+
+        openCLExecuteKernel(accum.clCxt, &imgproc_hough, "circlesAccumCenters", globalThreads, localThreads, args, -1, -1);        
+    }
+
+    int buildCentersList_gpu(const oclMat& accum, oclMat& centers, int threshold)
+    {
+        int totalCount = 0;
+        int err = CL_SUCCESS;
+        cl_mem counter = clCreateBuffer(accum.clCxt->impl->clContext,
+                                        CL_MEM_COPY_HOST_PTR,  
+                                        sizeof(int),
+                                        &totalCount,   
+                                        &err);
+        openCLSafeCall(err);        
+
+        const size_t blkSizeX = 32;
+        const size_t blkSizeY = 8;
+        size_t localThreads[3] = { blkSizeX, blkSizeY, 1 };        
+        
+        const size_t glbSizeX = (accum.cols - 2) % blkSizeX == 0 ? accum.cols - 2 : MUL_UP(accum.cols - 2, blkSizeX);
+        const size_t glbSizeY = (accum.rows - 2) % blkSizeY == 0 ? accum.rows - 2 : MUL_UP(accum.rows - 2, blkSizeY);
+        size_t globalThreads[3] = { glbSizeX, glbSizeY, 1 };
+
+        vector<pair<size_t , const void *> > args;
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&accum.data ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&accum.cols ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&accum.rows ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&accum.step ));
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&centers.data ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&threshold ));
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&counter ));
+
+        openCLExecuteKernel(accum.clCxt, &imgproc_hough, "buildCentersList", globalThreads, localThreads, args, -1, -1);
+        
+        openCLSafeCall(clEnqueueReadBuffer(accum.clCxt->impl->clCmdQueue, counter, CL_TRUE, 0, sizeof(int), &totalCount, 0, NULL, NULL));  
+        openCLSafeCall(clReleaseMemObject(counter));
+        
+        return totalCount;
+    }
+
+    int circlesAccumRadius_gpu(const oclMat& centers, int centersCount,
+                               const oclMat& list, int count,
+                               oclMat& circles, int maxCircles,
+                               float dp, int minRadius, int maxRadius, int threshold)
+    {
+        int totalCount = 0;
+        int err = CL_SUCCESS;
+        cl_mem counter = clCreateBuffer(circles.clCxt->impl->clContext,
+                                        CL_MEM_COPY_HOST_PTR,  
+                                        sizeof(int),
+                                        &totalCount,   
+                                        &err);
+        openCLSafeCall(err);        
+
+        const size_t blkSizeX = circles.clCxt->impl->maxWorkGroupSize; 
+        size_t localThreads[3] = { blkSizeX, 1, 1 };
+
+        const size_t glbSizeX = centersCount * blkSizeX;
+        size_t globalThreads[3] = { glbSizeX, 1, 1 };
+
+        const int histSize = maxRadius - minRadius + 1;
+        size_t smemSize = (histSize + 2) * sizeof(int);
+
+        vector<pair<size_t , const void *> > args;
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&centers.data ));
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&list.data ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&count ));
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&circles.data ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&maxCircles ));
+        args.push_back( make_pair( sizeof(cl_float), (void *)&dp ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&minRadius ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&maxRadius ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&histSize ));
+        args.push_back( make_pair( sizeof(cl_int)  , (void *)&threshold ));
+        args.push_back( make_pair( smemSize        , (void *)NULL ));
+        args.push_back( make_pair( sizeof(cl_mem)  , (void *)&counter ));
+
+        CV_Assert(circles.offset == 0);
+
+        openCLExecuteKernel(circles.clCxt, &imgproc_hough, "circlesAccumRadius", globalThreads, localThreads, args, -1, -1);        
+
+        openCLSafeCall(clEnqueueReadBuffer(circles.clCxt->impl->clCmdQueue, counter, CL_TRUE, 0, sizeof(int), &totalCount, 0, NULL, NULL));
+        
+        openCLSafeCall(clReleaseMemObject(counter));
+
+        totalCount = ::min(totalCount, maxCircles);
+
+        return totalCount;
+    }
+
+
+}}} // namespace cv { namespace ocl { namespace hough
+
+
+
+void cv::ocl::HoughCircles(const oclMat& src, oclMat& circles, int method, float dp, float minDist, int cannyThreshold, int votesThreshold, int minRadius, int maxRadius, int maxCircles)
+{
+    HoughCirclesBuf buf;
+    HoughCircles(src, circles, buf, method, dp, minDist, cannyThreshold, votesThreshold, minRadius, maxRadius, maxCircles);
+}
+
+void cv::ocl::HoughCircles(const oclMat& src, oclMat& circles, HoughCirclesBuf& buf, int method,
+                           float dp, float minDist, int cannyThreshold, int votesThreshold, int minRadius, int maxRadius, int maxCircles)
+{
+    CV_Assert(src.type() == CV_8UC1);
+    CV_Assert(src.cols < std::numeric_limits<unsigned short>::max());
+    CV_Assert(src.rows < std::numeric_limits<unsigned short>::max());
+    CV_Assert(method == CV_HOUGH_GRADIENT);
+    CV_Assert(dp > 0);
+    CV_Assert(minRadius > 0 && maxRadius > minRadius);
+    CV_Assert(cannyThreshold > 0);
+    CV_Assert(votesThreshold > 0);
+    CV_Assert(maxCircles > 0);
+    
+    const float idp = 1.0f / dp;
+
+    cv::ocl::Canny(src, buf.cannyBuf, buf.edges, std::max(cannyThreshold / 2, 1), cannyThreshold);
+
+    ensureSizeIsEnough(1, src.size().area(), CV_32SC1, buf.srcPoints);
+    const int pointsCount = hough::buildPointList_gpu(buf.edges, buf.srcPoints);
+    if (pointsCount == 0)
+    {
+        circles.release();
+        return;
+    }
+
+    ensureSizeIsEnough(cvCeil(src.rows * idp) + 2, cvCeil(src.cols * idp) + 2, CV_32SC1, buf.accum);
+    buf.accum.setTo(Scalar::all(0));
+
+    hough::circlesAccumCenters_gpu(buf.srcPoints, pointsCount, buf.cannyBuf.dx, buf.cannyBuf.dy, buf.accum, minRadius, maxRadius, idp);
+
+    ensureSizeIsEnough(1, src.size().area(), CV_32SC1, buf.centers);    
+    int centersCount = hough::buildCentersList_gpu(buf.accum, buf.centers, votesThreshold);
+    if (centersCount == 0)
+    {
+        circles.release();
+        return;
+    }
+
+    if (minDist > 1)
+    {
+        cv::AutoBuffer<unsigned int> oldBuf_(centersCount);
+        cv::AutoBuffer<unsigned int> newBuf_(centersCount);
+        int newCount = 0;
+
+        unsigned int* oldBuf = oldBuf_;
+        unsigned int* newBuf = newBuf_;
+
+        openCLSafeCall(clEnqueueReadBuffer(buf.centers.clCxt->impl->clCmdQueue,
+                                           (cl_mem)buf.centers.data,
+                                           CL_TRUE,
+                                           0,
+                                           centersCount * sizeof(unsigned int),
+                                           oldBuf,
+                                           0,
+                                           NULL,
+                                           NULL));  
+
+        
+        const int cellSize = cvRound(minDist);
+        const int gridWidth = (src.cols + cellSize - 1) / cellSize;
+        const int gridHeight = (src.rows + cellSize - 1) / cellSize;
+
+        std::vector< std::vector<unsigned int> > grid(gridWidth * gridHeight);
+
+        const float minDist2 = minDist * minDist;
+
+        for (int i = 0; i < centersCount; ++i)
+        {
+            unsigned int p = oldBuf[i];
+            const int px = p & 0xFFFF;
+            const int py = (p >> 16) & 0xFFFF;
+
+            bool good = true;
+
+            int xCell = static_cast<int>(px / cellSize);
+            int yCell = static_cast<int>(py / cellSize);
+
+            int x1 = xCell - 1;
+            int y1 = yCell - 1;
+            int x2 = xCell + 1;
+            int y2 = yCell + 1;
+
+            // boundary check
+            x1 = std::max(0, x1);
+            y1 = std::max(0, y1);
+            x2 = std::min(gridWidth - 1, x2);
+            y2 = std::min(gridHeight - 1, y2);
+
+            for (int yy = y1; yy <= y2; ++yy)
+            {
+                for (int xx = x1; xx <= x2; ++xx)
+                {
+                    vector<unsigned int>& m = grid[yy * gridWidth + xx];
+
+                    for(size_t j = 0; j < m.size(); ++j)
+                    {
+                        const int val = m[j];
+                        const int jx = val & 0xFFFF;
+                        const int jy = (val >> 16) & 0xFFFF;
+                        
+                        float dx = (float)(px - jx);
+                        float dy = (float)(py - jy);
+
+                        if (dx * dx + dy * dy < minDist2)
+                        {
+                            good = false;
+                            goto break_out;
+                        }
+                    }
+                }
+            }
+
+            break_out:
+
+            if(good)
+            {
+                grid[yCell * gridWidth + xCell].push_back(p);
+                newBuf[newCount++] = p;
+            }
+        }
+
+        openCLSafeCall(clEnqueueWriteBuffer(buf.centers.clCxt->impl->clCmdQueue,
+                                            (cl_mem)buf.centers.data,
+                                            CL_TRUE,
+                                            0,
+                                            newCount * sizeof(unsigned int),
+                                            newBuf,
+                                            0,
+                                            0,
+                                            0));     
+        centersCount = newCount;
+    }
+
+    ensureSizeIsEnough(1, maxCircles, CV_32FC3, circles);
+
+    const int circlesCount = hough::circlesAccumRadius_gpu(buf.centers, centersCount,
+                                                           buf.srcPoints, pointsCount,
+                                                           circles, maxCircles,
+                                                           dp, minRadius, maxRadius, votesThreshold);
+
+    if (circlesCount > 0)
+        circles.cols = circlesCount;
+    else
+        circles.release();
+}
+
+void cv::ocl::HoughCirclesDownload(const oclMat& d_circles, cv::OutputArray h_circles_)
+{
+    // FIX ME: garbage values are copied!
+    CV_Error(CV_StsNotImplemented, "HoughCirclesDownload is not implemented");
+     
+    if (d_circles.empty())
+    {
+        h_circles_.release();
+        return;
+    }
+
+    CV_Assert(d_circles.rows == 1 && d_circles.type() == CV_32FC3);
+
+    h_circles_.create(1, d_circles.cols, CV_32FC3);
+    Mat h_circles = h_circles_.getMat();
+    d_circles.download(h_circles);
+}
+
+#endif /* !defined (HAVE_OPENCL) */

modules/ocl/src/kernels/imgproc_hough.cl

@@ -0,0 +1,280 @@
+/*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) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage 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 bpied warranties, including, but not limited to, the bpied
+// 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*/
+
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+
+////////////////////////////////////////////////////////////////////////
+// buildPointList
+
+#define PIXELS_PER_THREAD 16
+
+// TODO: add offset to support ROI
+__kernel void buildPointList(__global const uchar* src,
+                             int cols,
+                             int rows,
+                             int step,
+                             __global unsigned int* list,
+                             __global int* counter)
+{
+    __local unsigned int s_queues[4][32 * PIXELS_PER_THREAD];
+    __local int s_qsize[4];
+    __local int s_globStart[4];
+
+    const int x = get_group_id(0) * get_local_size(0) * PIXELS_PER_THREAD + get_local_id(0);
+    const int y = get_global_id(1);
+
+    if (get_local_id(0) == 0)
+        s_qsize[get_local_id(1)] = 0;
+    barrier(CLK_LOCAL_MEM_FENCE);
+        
+    if (y < rows)
+    {
+        // fill the queue
+        __global const uchar* srcRow = &src[y * step];
+        for (int i = 0, xx = x; i < PIXELS_PER_THREAD && xx < cols; ++i, xx += get_local_size(0))
+        {
+            if (srcRow[xx])
+            {
+                const unsigned int val = (y << 16) | xx;
+                const int qidx = atomic_add(&s_qsize[get_local_id(1)], 1);
+                s_queues[get_local_id(1)][qidx] = val;
+            }
+        }
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    // let one work-item reserve the space required in the global list
+    if (get_local_id(0) == 0 && get_local_id(1) == 0)
+    {
+        // find how many items are stored in each list
+        int totalSize = 0;
+        for (int i = 0; i < get_local_size(1); ++i)
+        {
+            s_globStart[i] = totalSize;
+            totalSize += s_qsize[i];
+        }
+
+        // calculate the offset in the global list
+        const int globalOffset = atomic_add(counter, totalSize);
+        for (int i = 0; i < get_local_size(1); ++i)
+            s_globStart[i] += globalOffset;
+    }
+
+    barrier(CLK_GLOBAL_MEM_FENCE);
+    
+    // copy local queues to global queue
+    const int qsize = s_qsize[get_local_id(1)];
+    int gidx = s_globStart[get_local_id(1)] + get_local_id(0);
+    for(int i = get_local_id(0); i < qsize; i += get_local_size(0), gidx += get_local_size(0))
+        list[gidx] = s_queues[get_local_id(1)][i];
+}
+
+////////////////////////////////////////////////////////////////////////
+// circlesAccumCenters
+
+// TODO: add offset to support ROI
+__kernel void circlesAccumCenters(__global const unsigned int* list,
+                                  const int count,
+                                  __global const int* dx,
+                                  const int dxStep,
+                                  __global const int* dy,
+                                  const int dyStep,
+                                  __global int* accum,
+                                  const int accumStep,
+                                  const int width,
+                                  const int height,
+                                  const int minRadius,
+                                  const int maxRadius,
+                                  const float idp)
+{
+    const int dxStepInPixel    = dxStep    / sizeof(int);
+    const int dyStepInPixel    = dyStep    / sizeof(int);
+    const int accumStepInPixel = accumStep / sizeof(int);
+    
+    const int SHIFT = 10;
+    const int ONE = 1 << SHIFT;
+
+    // const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+    const int wid = get_global_id(0);
+
+    if (wid >= count)
+        return;
+
+    const unsigned int val = list[wid];
+
+    const int x = (val & 0xFFFF);
+    const int y = (val >> 16) & 0xFFFF;
+
+    const int vx = dx[mad24(y, dxStepInPixel, x)];
+    const int vy = dy[mad24(y, dyStepInPixel, x)];
+
+    if (vx == 0 && vy == 0)
+        return;
+
+    const float mag = sqrt(convert_float(vx * vx + vy * vy));
+
+    const int x0 = convert_int_rte((x * idp) * ONE);
+    const int y0 = convert_int_rte((y * idp) * ONE);
+
+    int sx = convert_int_rte((vx * idp) * ONE / mag);
+    int sy = convert_int_rte((vy * idp) * ONE / mag);
+
+    // Step from minRadius to maxRadius in both directions of the gradient
+    for (int k1 = 0; k1 < 2; ++k1)
+    {
+        int x1 = x0 + minRadius * sx;
+        int y1 = y0 + minRadius * sy;
+
+        for (int r = minRadius; r <= maxRadius; x1 += sx, y1 += sy, ++r)
+        {
+            const int x2 = x1 >> SHIFT;
+            const int y2 = y1 >> SHIFT;
+
+            if (x2 < 0 || x2 >= width || y2 < 0 || y2 >= height)
+                break;
+
+            atomic_add(&accum[mad24(y2+1, accumStepInPixel, x2+1)], 1);
+        }
+
+        sx = -sx;
+        sy = -sy;
+    }
+}
+
+// ////////////////////////////////////////////////////////////////////////
+// // buildCentersList
+
+// TODO: add offset to support ROI
+__kernel void buildCentersList(__global const int* accum,
+                               const int accumCols,
+                               const int accumRows,
+                               const int accumStep,
+                               __global unsigned int* centers,
+                               const int threshold,
+                               __global int* counter)
+{
+    const int accumStepInPixel = accumStep/sizeof(int);
+    
+    const int x = get_global_id(0);
+    const int y = get_global_id(1);
+
+    if (x < accumCols - 2 && y < accumRows - 2)
+    {
+        const int top    = accum[mad24(y,     accumStepInPixel, x + 1)];
+
+        const int left   = accum[mad24(y + 1, accumStepInPixel, x)];
+        const int cur    = accum[mad24(y + 1, accumStepInPixel, x + 1)];
+        const int right  = accum[mad24(y + 1, accumStepInPixel, x + 2)];
+        
+        const int bottom = accum[mad24(y + 2, accumStepInPixel, x + 1)];;
+
+        if (cur > threshold && cur > top && cur >= bottom && cur >  left && cur >= right)
+        {
+            const unsigned int val = (y << 16) | x;
+            const int idx = atomic_add(counter, 1);
+            centers[idx] = val;
+        }
+    }
+}
+
+
+// ////////////////////////////////////////////////////////////////////////
+// // circlesAccumRadius
+
+// TODO: add offset to support ROI
+__kernel void circlesAccumRadius(__global const unsigned int* centers,
+                                 __global const unsigned int* list, const int count,
+                                 __global float4* circles, const int maxCircles,
+                                 const float dp,
+                                 const int minRadius, const int maxRadius,
+                                 const int histSize,
+                                 const int threshold,
+                                 __local int* smem,
+                                 __global int* counter)
+{
+    for (int i = get_local_id(0); i < histSize + 2; i += get_local_size(0))
+        smem[i] = 0;
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    unsigned int val = centers[get_group_id(0)];
+
+    float cx = convert_float(val & 0xFFFF);
+    float cy = convert_float((val >> 16) & 0xFFFF);
+
+    cx = (cx + 0.5f) * dp;
+    cy = (cy + 0.5f) * dp;
+
+    for (int i = get_local_id(0); i < count; i += get_local_size(0))
+    {
+        val = list[i];
+
+        const int x = (val & 0xFFFF);
+        const int y = (val >> 16) & 0xFFFF;
+
+        const float rad = sqrt((cx - x) * (cx - x) + (cy - y) * (cy - y));
+        if (rad >= minRadius && rad <= maxRadius)
+        {
+            const int r = convert_int_rte(rad - minRadius);
+
+            atomic_add(&smem[r + 1], 1);
+        }
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    for (int i = get_local_id(0); i < histSize; i += get_local_size(0))
+    {
+        const int curVotes = smem[i + 1];
+
+        if (curVotes >= threshold && curVotes > smem[i] && curVotes >= smem[i + 2])
+            
+        {
+            const int ind = atomic_add(counter, 1);
+            if (ind < maxCircles)
+            {
+                circles[ind] = (float4)(cx, cy, convert_float(i + minRadius), 0.0f);
+            }
+        }
+    }
+}

modules/ocl/test/test_hough.cpp

@@ -0,0 +1,112 @@
+/*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) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2008-2011, Willow Garage 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 Intel Corporation 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*/
+
+#include "precomp.hpp"
+
+#ifdef HAVE_OPENCL
+
+///////////////////////////////////////////////////////////////////////////////////////////////////////
+// HoughCircles
+
+PARAM_TEST_CASE(HoughCircles, cv::Size)
+{
+    static void drawCircles(cv::Mat& dst, const std::vector<cv::Vec3f>& circles, bool fill)
+    {
+        dst.setTo(cv::Scalar::all(0));
+
+        for (size_t i = 0; i < circles.size(); ++i)
+            cv::circle(dst, cv::Point2f(circles[i][0], circles[i][1]), (int)circles[i][2], cv::Scalar::all(255), fill ? -1 : 1);
+    }
+};
+
+TEST_P(HoughCircles, Accuracy)
+{
+    const cv::Size size = GET_PARAM(0);
+
+    const float dp = 2.0f;
+    const float minDist = 10.0f;
+    const int minRadius = 10;
+    const int maxRadius = 20;
+    const int cannyThreshold = 100;
+    const int votesThreshold = 15;
+
+    std::vector<cv::Vec3f> circles_gold(4);
+    circles_gold[0] = cv::Vec3i(20, 20, minRadius);
+    circles_gold[1] = cv::Vec3i(90, 87, minRadius + 3);
+    circles_gold[2] = cv::Vec3i(30, 70, minRadius + 8);
+    circles_gold[3] = cv::Vec3i(80, 10, maxRadius);
+
+    cv::Mat src(size, CV_8UC1);
+    drawCircles(src, circles_gold, true);
+    cv::ocl::oclMat d_src(src);
+
+    cv::ocl::oclMat d_circles;
+    cv::ocl::HoughCircles(d_src, d_circles, CV_HOUGH_GRADIENT, dp, minDist, cannyThreshold, votesThreshold, minRadius, maxRadius);
+    ASSERT_TRUE(d_circles.rows > 0);
+
+    cv::Mat circles;
+    d_circles.download(circles);
+
+    for (int i = 0; i < circles.cols; ++i)
+    {
+        cv::Vec3f cur = circles.at<cv::Vec3f>(i);
+
+        bool found = false;
+
+        for (size_t j = 0; j < circles_gold.size(); ++j)
+        {
+            cv::Vec3f gold = circles_gold[j];
+
+            if (std::fabs(cur[0] - gold[0]) < minDist && std::fabs(cur[1] - gold[1]) < minDist && std::fabs(cur[2] - gold[2]) < minDist)
+            {
+                found = true;
+                break;
+            }
+        }
+
+        ASSERT_TRUE(found);
+    }
+}
+
+INSTANTIATE_TEST_CASE_P(Hough, HoughCircles, DIFFERENT_SIZES);
+
+#endif // HAVE_OPENCL