added gpu BM optical flow implementation

modules/gpu/include/opencv2/gpu/gpu.hpp

@@ -2074,6 +2074,24 @@ private:
 };
 
 
+//! Calculates optical flow for 2 images using block matching algorithm */
+CV_EXPORTS void calcOpticalFlowBM(const GpuMat& prev, const GpuMat& curr,
+                                  Size block_size, Size shift_size, Size max_range, bool use_previous,
+                                  GpuMat& velx, GpuMat& vely, GpuMat& buf,
+                                  Stream& stream = Stream::Null());
+
+class CV_EXPORTS FastOpticalFlowBM
+{
+public:
+    void operator ()(const GpuMat& I0, const GpuMat& I1, GpuMat& flowx, GpuMat& flowy, int search_window = 21, int block_window = 7, Stream& s = Stream::Null());
+
+private:
+    GpuMat buffer;
+    GpuMat extended_I0;
+    GpuMat extended_I1;
+};
+
+
 //! Interpolate frames (images) using provided optical flow (displacement field).
 //! frame0   - frame 0 (32-bit floating point images, single channel)
 //! frame1   - frame 1 (the same type and size)

modules/gpu/perf/perf_video.cpp

@@ -444,6 +444,123 @@ PERF_TEST_P(ImagePair, Video_OpticalFlowDual_TVL1,
     }
 }
 
+//////////////////////////////////////////////////////
+// OpticalFlowBM
+
+void calcOpticalFlowBM(const cv::Mat& prev, const cv::Mat& curr,
+                       cv::Size bSize, cv::Size shiftSize, cv::Size maxRange, int usePrevious,
+                       cv::Mat& velx, cv::Mat& vely)
+{
+    cv::Size sz((curr.cols - bSize.width + shiftSize.width)/shiftSize.width, (curr.rows - bSize.height + shiftSize.height)/shiftSize.height);
+
+    velx.create(sz, CV_32FC1);
+    vely.create(sz, CV_32FC1);
+
+    CvMat cvprev = prev;
+    CvMat cvcurr = curr;
+
+    CvMat cvvelx = velx;
+    CvMat cvvely = vely;
+
+    cvCalcOpticalFlowBM(&cvprev, &cvcurr, bSize, shiftSize, maxRange, usePrevious, &cvvelx, &cvvely);
+}
+
+PERF_TEST_P(ImagePair, Video_OpticalFlowBM,
+    Values<pair_string>(make_pair("gpu/opticalflow/frame0.png", "gpu/opticalflow/frame1.png")))
+{
+    declare.time(400);
+
+    cv::Mat frame0 = readImage(GetParam().first, cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(frame0.empty());
+
+    cv::Mat frame1 = readImage(GetParam().second, cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(frame1.empty());
+
+    cv::Size block_size(16, 16);
+    cv::Size shift_size(1, 1);
+    cv::Size max_range(16, 16);
+
+    if (PERF_RUN_GPU())
+    {
+        cv::gpu::GpuMat d_frame0(frame0);
+        cv::gpu::GpuMat d_frame1(frame1);
+        cv::gpu::GpuMat d_velx, d_vely, buf;
+
+        cv::gpu::calcOpticalFlowBM(d_frame0, d_frame1, block_size, shift_size, max_range, false, d_velx, d_vely, buf);
+
+        TEST_CYCLE()
+        {
+            cv::gpu::calcOpticalFlowBM(d_frame0, d_frame1, block_size, shift_size, max_range, false, d_velx, d_vely, buf);
+        }
+
+        GPU_SANITY_CHECK(d_velx);
+        GPU_SANITY_CHECK(d_vely);
+    }
+    else
+    {
+        cv::Mat velx, vely;
+
+        calcOpticalFlowBM(frame0, frame1, block_size, shift_size, max_range, false, velx, vely);
+
+        TEST_CYCLE()
+        {
+            calcOpticalFlowBM(frame0, frame1, block_size, shift_size, max_range, false, velx, vely);
+        }
+
+        CPU_SANITY_CHECK(velx);
+        CPU_SANITY_CHECK(vely);
+    }
+}
+
+PERF_TEST_P(ImagePair, Video_FastOpticalFlowBM,
+    Values<pair_string>(make_pair("gpu/opticalflow/frame0.png", "gpu/opticalflow/frame1.png")))
+{
+    declare.time(400);
+
+    cv::Mat frame0 = readImage(GetParam().first, cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(frame0.empty());
+
+    cv::Mat frame1 = readImage(GetParam().second, cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(frame1.empty());
+
+    cv::Size block_size(16, 16);
+    cv::Size shift_size(1, 1);
+    cv::Size max_range(16, 16);
+
+    if (PERF_RUN_GPU())
+    {
+        cv::gpu::GpuMat d_frame0(frame0);
+        cv::gpu::GpuMat d_frame1(frame1);
+        cv::gpu::GpuMat d_velx, d_vely;
+
+        cv::gpu::FastOpticalFlowBM fastBM;
+
+        fastBM(d_frame0, d_frame1, d_velx, d_vely, max_range.width, block_size.width);
+
+        TEST_CYCLE()
+        {
+            fastBM(d_frame0, d_frame1, d_velx, d_vely, max_range.width, block_size.width);
+        }
+
+        GPU_SANITY_CHECK(d_velx);
+        GPU_SANITY_CHECK(d_vely);
+    }
+    else
+    {
+        cv::Mat velx, vely;
+
+        calcOpticalFlowBM(frame0, frame1, block_size, shift_size, max_range, false, velx, vely);
+
+        TEST_CYCLE()
+        {
+            calcOpticalFlowBM(frame0, frame1, block_size, shift_size, max_range, false, velx, vely);
+        }
+
+        CPU_SANITY_CHECK(velx);
+        CPU_SANITY_CHECK(vely);
+    }
+}
+
 //////////////////////////////////////////////////////
 // FGDStatModel
 

modules/gpu/src/cuda/optflowbm.cu

@@ -0,0 +1,414 @@
+/*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*/
+
+#if !defined CUDA_DISABLER
+
+#include "opencv2/gpu/device/common.hpp"
+#include "opencv2/gpu/device/limits.hpp"
+#include "opencv2/gpu/device/functional.hpp"
+#include "opencv2/gpu/device/reduce.hpp"
+
+using namespace cv::gpu;
+using namespace cv::gpu::device;
+
+namespace optflowbm
+{
+    texture<uchar, cudaTextureType2D, cudaReadModeElementType> tex_prev(false, cudaFilterModePoint, cudaAddressModeClamp);
+    texture<uchar, cudaTextureType2D, cudaReadModeElementType> tex_curr(false, cudaFilterModePoint, cudaAddressModeClamp);
+
+    __device__ int cmpBlocks(int X1, int Y1, int X2, int Y2, int2 blockSize)
+    {
+        int s = 0;
+
+        for (int y = 0; y < blockSize.y; ++y)
+        {
+            for (int x = 0; x < blockSize.x; ++x)
+                s += ::abs(tex2D(tex_prev, X1 + x, Y1 + y) - tex2D(tex_curr, X2 + x, Y2 + y));
+        }
+
+        return s;
+    }
+
+    __global__ void calcOptFlowBM(PtrStepSzf velx, PtrStepf vely, const int2 blockSize, const int2 shiftSize, const bool usePrevious,
+                                  const int maxX, const int maxY, const int acceptLevel, const int escapeLevel,
+                                  const short2* ss, const int ssCount)
+    {
+        const int j = blockIdx.x * blockDim.x + threadIdx.x;
+        const int i = blockIdx.y * blockDim.y + threadIdx.y;
+
+        if (i >= velx.rows || j >= velx.cols)
+            return;
+
+        const int X1 = j * shiftSize.x;
+        const int Y1 = i * shiftSize.y;
+
+        const int offX = usePrevious ? __float2int_rn(velx(i, j)) : 0;
+        const int offY = usePrevious ? __float2int_rn(vely(i, j)) : 0;
+
+        int X2 = X1 + offX;
+        int Y2 = Y1 + offY;
+
+        int dist = numeric_limits<int>::max();
+
+        if (0 <= X2 && X2 <= maxX && 0 <= Y2 && Y2 <= maxY)
+            dist = cmpBlocks(X1, Y1, X2, Y2, blockSize);
+
+        int countMin = 1;
+        int sumx = offX;
+        int sumy = offY;
+
+        if (dist > acceptLevel)
+        {
+            // do brute-force search
+            for (int k = 0; k < ssCount; ++k)
+            {
+                const short2 ssVal = ss[k];
+
+                const int dx = offX + ssVal.x;
+                const int dy = offY + ssVal.y;
+
+                X2 = X1 + dx;
+                Y2 = Y1 + dy;
+
+                if (0 <= X2 && X2 <= maxX && 0 <= Y2 && Y2 <= maxY)
+                {
+                    const int tmpDist = cmpBlocks(X1, Y1, X2, Y2, blockSize);
+                    if (tmpDist < acceptLevel)
+                    {
+                        sumx = dx;
+                        sumy = dy;
+                        countMin = 1;
+                        break;
+                    }
+
+                    if (tmpDist < dist)
+                    {
+                        dist = tmpDist;
+                        sumx = dx;
+                        sumy = dy;
+                        countMin = 1;
+                    }
+                    else if (tmpDist == dist)
+                    {
+                        sumx += dx;
+                        sumy += dy;
+                        countMin++;
+                    }
+                }
+            }
+
+            if (dist > escapeLevel)
+            {
+                sumx = offX;
+                sumy = offY;
+                countMin = 1;
+            }
+        }
+
+        velx(i, j) = static_cast<float>(sumx) / countMin;
+        vely(i, j) = static_cast<float>(sumy) / countMin;
+    }
+
+    void calc(PtrStepSzb prev, PtrStepSzb curr, PtrStepSzf velx, PtrStepSzf vely, int2 blockSize, int2 shiftSize, bool usePrevious,
+              int maxX, int maxY, int acceptLevel, int escapeLevel, const short2* ss, int ssCount, cudaStream_t stream)
+    {
+        bindTexture(&tex_prev, prev);
+        bindTexture(&tex_curr, curr);
+
+        const dim3 block(32, 8);
+        const dim3 grid(divUp(velx.cols, block.x), divUp(vely.rows, block.y));
+
+        calcOptFlowBM<<<grid, block, 0, stream>>>(velx, vely, blockSize, shiftSize, usePrevious,
+                                                  maxX, maxY, acceptLevel,  escapeLevel, ss, ssCount);
+        cudaSafeCall( cudaGetLastError() );
+
+        if (stream == 0)
+            cudaSafeCall( cudaDeviceSynchronize() );
+    }
+}
+
+/////////////////////////////////////////////////////////
+// Fast approximate version
+
+namespace optflowbm_fast
+{
+    enum
+    {
+        CTA_SIZE = 128,
+
+        TILE_COLS = 128,
+        TILE_ROWS = 32,
+
+        STRIDE = CTA_SIZE
+    };
+
+    template <typename T> __device__ __forceinline__ int calcDist(T a, T b)
+    {
+        return ::abs(a - b);
+    }
+
+    template <class T> struct FastOptFlowBM
+    {
+
+        int search_radius;
+        int block_radius;
+
+        int search_window;
+        int block_window;
+
+        PtrStepSz<T> I0;
+        PtrStep<T> I1;
+
+        mutable PtrStepi buffer;
+
+        FastOptFlowBM(int search_window_, int block_window_,
+                      PtrStepSz<T> I0_, PtrStepSz<T> I1_,
+                      PtrStepi buffer_) :
+            search_radius(search_window_ / 2), block_radius(block_window_ / 2),
+            search_window(search_window_), block_window(block_window_),
+            I0(I0_), I1(I1_),
+            buffer(buffer_)
+        {
+        }
+
+        __device__ __forceinline__ void initSums_BruteForce(int i, int j, int* dist_sums, PtrStepi& col_sums, PtrStepi& up_col_sums) const
+        {
+            for (int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
+            {
+                dist_sums[index] = 0;
+
+                for (int tx = 0; tx < block_window; ++tx)
+                    col_sums(tx, index) = 0;
+
+                int y = index / search_window;
+                int x = index - y * search_window;
+
+                int ay = i;
+                int ax = j;
+
+                int by = i + y - search_radius;
+                int bx = j + x - search_radius;
+
+                for (int tx = -block_radius; tx <= block_radius; ++tx)
+                {
+                    int col_sum = 0;
+                    for (int ty = -block_radius; ty <= block_radius; ++ty)
+                    {
+                        int dist = calcDist(I0(ay + ty, ax + tx), I1(by + ty, bx + tx));
+
+                        dist_sums[index] += dist;
+                        col_sum += dist;
+                    }
+
+                    col_sums(tx + block_radius, index) = col_sum;
+                }
+
+                up_col_sums(j, index) = col_sums(block_window - 1, index);
+            }
+        }
+
+        __device__ __forceinline__ void shiftRight_FirstRow(int i, int j, int first, int* dist_sums, PtrStepi& col_sums, PtrStepi& up_col_sums) const
+        {
+            for (int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
+            {
+                int y = index / search_window;
+                int x = index - y * search_window;
+
+                int ay = i;
+                int ax = j + block_radius;
+
+                int by = i + y - search_radius;
+                int bx = j + x - search_radius + block_radius;
+
+                int col_sum = 0;
+
+                for (int ty = -block_radius; ty <= block_radius; ++ty)
+                    col_sum += calcDist(I0(ay + ty, ax), I1(by + ty, bx));
+
+                dist_sums[index] += col_sum - col_sums(first, index);
+
+                col_sums(first, index) = col_sum;
+                up_col_sums(j, index) = col_sum;
+            }
+        }
+
+        __device__ __forceinline__ void shiftRight_UpSums(int i, int j, int first, int* dist_sums, PtrStepi& col_sums, PtrStepi& up_col_sums) const
+        {
+            int ay = i;
+            int ax = j + block_radius;
+
+            T a_up   = I0(ay - block_radius - 1, ax);
+            T a_down = I0(ay + block_radius, ax);
+
+            for(int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
+            {
+                int y = index / search_window;
+                int x = index - y * search_window;
+
+                int by = i + y - search_radius;
+                int bx = j + x - search_radius + block_radius;
+
+                T b_up   = I1(by - block_radius - 1, bx);
+                T b_down = I1(by + block_radius, bx);
+
+                int col_sum = up_col_sums(j, index) + calcDist(a_down, b_down) - calcDist(a_up, b_up);
+
+                dist_sums[index] += col_sum  - col_sums(first, index);
+                col_sums(first, index) = col_sum;
+                up_col_sums(j, index) = col_sum;
+            }
+        }
+
+        __device__ __forceinline__ void convolve_window(int i, int j, const int* dist_sums, float& velx, float& vely) const
+        {
+            int bestDist = numeric_limits<int>::max();
+            int bestInd = -1;
+
+            for (int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
+            {
+                int curDist = dist_sums[index];
+                if (curDist < bestDist)
+                {
+                    bestDist = curDist;
+                    bestInd = index;
+                }
+            }
+
+            __shared__ int cta_dist_buffer[CTA_SIZE];
+            __shared__ int cta_ind_buffer[CTA_SIZE];
+
+            reduceKeyVal<CTA_SIZE>(cta_dist_buffer, bestDist, cta_ind_buffer, bestInd, threadIdx.x, less<int>());
+
+            if (threadIdx.x == 0)
+            {
+                int y = bestInd / search_window;
+                int x = bestInd - y * search_window;
+
+                velx = x - search_radius;
+                vely = y - search_radius;
+            }
+        }
+
+        __device__ __forceinline__ void operator()(PtrStepf velx, PtrStepf vely) const
+        {
+            int tbx = blockIdx.x * TILE_COLS;
+            int tby = blockIdx.y * TILE_ROWS;
+
+            int tex = ::min(tbx + TILE_COLS, I0.cols);
+            int tey = ::min(tby + TILE_ROWS, I0.rows);
+
+            PtrStepi col_sums;
+            col_sums.data = buffer.ptr(I0.cols + blockIdx.x * block_window) + blockIdx.y * search_window * search_window;
+            col_sums.step = buffer.step;
+
+            PtrStepi up_col_sums;
+            up_col_sums.data = buffer.data + blockIdx.y * search_window * search_window;
+            up_col_sums.step = buffer.step;
+
+            extern __shared__ int dist_sums[]; //search_window * search_window
+
+            int first = 0;
+
+            for (int i = tby; i < tey; ++i)
+            {
+                for (int j = tbx; j < tex; ++j)
+                {
+                    __syncthreads();
+
+                    if (j == tbx)
+                    {
+                        initSums_BruteForce(i, j, dist_sums, col_sums, up_col_sums);
+                        first = 0;
+                    }
+                    else
+                    {
+                        if (i == tby)
+                          shiftRight_FirstRow(i, j, first, dist_sums, col_sums, up_col_sums);
+                        else
+                          shiftRight_UpSums(i, j, first, dist_sums, col_sums, up_col_sums);
+
+                        first = (first + 1) % block_window;
+                    }
+
+                    __syncthreads();
+
+                    convolve_window(i, j, dist_sums, velx(i, j), vely(i, j));
+                }
+            }
+        }
+
+    };
+
+    template<typename T> __global__ void optflowbm_fast_kernel(const FastOptFlowBM<T> fbm, PtrStepf velx, PtrStepf vely)
+    {
+        fbm(velx, vely);
+    }
+
+    void get_buffer_size(int src_cols, int src_rows, int search_window, int block_window, int& buffer_cols, int& buffer_rows)
+    {
+        dim3 grid(divUp(src_cols, TILE_COLS), divUp(src_rows, TILE_ROWS));
+
+        buffer_cols = search_window * search_window * grid.y;
+        buffer_rows = src_cols + block_window * grid.x;
+    }
+
+    template <typename T>
+    void calc(PtrStepSzb I0, PtrStepSzb I1, PtrStepSzf velx, PtrStepSzf vely, PtrStepi buffer, int search_window, int block_window, cudaStream_t stream)
+    {
+        FastOptFlowBM<T> fbm(search_window, block_window, I0, I1, buffer);
+
+        dim3 block(CTA_SIZE, 1);
+        dim3 grid(divUp(I0.cols, TILE_COLS), divUp(I0.rows, TILE_ROWS));
+
+        size_t smem = search_window * search_window * sizeof(int);
+
+        optflowbm_fast_kernel<<<grid, block, smem, stream>>>(fbm, velx, vely);
+        cudaSafeCall ( cudaGetLastError () );
+
+        if (stream == 0)
+            cudaSafeCall( cudaDeviceSynchronize() );
+    }
+
+    template void calc<uchar>(PtrStepSzb I0, PtrStepSzb I1, PtrStepSzf velx, PtrStepSzf vely, PtrStepi buffer, int search_window, int block_window, cudaStream_t stream);
+}
+
+#endif // !defined CUDA_DISABLER

modules/gpu/src/optflowbm.cpp

@@ -0,0 +1,243 @@
+/*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::gpu;
+
+#if !defined HAVE_CUDA || defined(CUDA_DISABLER)
+
+void cv::gpu::calcOpticalFlowBM(const GpuMat&, const GpuMat&, Size, Size, Size, bool, GpuMat&, GpuMat&, GpuMat&, Stream&) { throw_nogpu(); }
+
+void cv::gpu::FastOpticalFlowBM::operator ()(const GpuMat&, const GpuMat&, GpuMat&, GpuMat&, int, int, Stream&) { throw_nogpu(); }
+
+#else // HAVE_CUDA
+
+namespace optflowbm
+{
+    void calc(PtrStepSzb prev, PtrStepSzb curr, PtrStepSzf velx, PtrStepSzf vely, int2 blockSize, int2 shiftSize, bool usePrevious,
+              int maxX, int maxY, int acceptLevel, int escapeLevel, const short2* ss, int ssCount, cudaStream_t stream);
+}
+
+void cv::gpu::calcOpticalFlowBM(const GpuMat& prev, const GpuMat& curr, Size blockSize, Size shiftSize, Size maxRange, bool usePrevious, GpuMat& velx, GpuMat& vely, GpuMat& buf, Stream& st)
+{
+    CV_Assert( prev.type() == CV_8UC1 );
+    CV_Assert( curr.size() == prev.size() && curr.type() == prev.type() );
+
+    const Size velSize((prev.cols - blockSize.width + shiftSize.width) / shiftSize.width,
+                       (prev.rows - blockSize.height + shiftSize.height) / shiftSize.height);
+
+    velx.create(velSize, CV_32FC1);
+    vely.create(velSize, CV_32FC1);
+
+    // scanning scheme coordinates
+    vector<short2> ss((2 * maxRange.width + 1) * (2 * maxRange.height + 1));
+    int ssCount = 0;
+
+    // Calculate scanning scheme
+    const int minCount = std::min(maxRange.width, maxRange.height);
+
+    // use spiral search pattern
+    //
+    //     9 10 11 12
+    //     8  1  2 13
+    //     7  *  3 14
+    //     6  5  4 15
+    //... 20 19 18 17
+    //
+
+    for (int i = 0; i < minCount; ++i)
+    {
+        // four cycles along sides
+        int x = -i - 1, y = x;
+
+        // upper side
+        for (int j = -i; j <= i + 1; ++j, ++ssCount)
+        {
+            ss[ssCount].x = ++x;
+            ss[ssCount].y = y;
+        }
+
+        // right side
+        for (int j = -i; j <= i + 1; ++j, ++ssCount)
+        {
+            ss[ssCount].x = x;
+            ss[ssCount].y = ++y;
+        }
+
+        // bottom side
+        for (int j = -i; j <= i + 1; ++j, ++ssCount)
+        {
+            ss[ssCount].x = --x;
+            ss[ssCount].y = y;
+        }
+
+        // left side
+        for (int j = -i; j <= i + 1; ++j, ++ssCount)
+        {
+            ss[ssCount].x = x;
+            ss[ssCount].y = --y;
+        }
+    }
+
+    // the rest part
+    if (maxRange.width < maxRange.height)
+    {
+        const int xleft = -minCount;
+
+        // cycle by neighbor rings
+        for (int i = minCount; i < maxRange.height; ++i)
+        {
+            // two cycles by x
+            int y = -(i + 1);
+            int x = xleft;
+
+            // upper side
+            for (int j = -maxRange.width; j <= maxRange.width; ++j, ++ssCount, ++x)
+            {
+                ss[ssCount].x = x;
+                ss[ssCount].y = y;
+            }
+
+            x = xleft;
+            y = -y;
+
+            // bottom side
+            for (int j = -maxRange.width; j <= maxRange.width; ++j, ++ssCount, ++x)
+            {
+                ss[ssCount].x = x;
+                ss[ssCount].y = y;
+            }
+        }
+    }
+    else if (maxRange.width > maxRange.height)
+    {
+        const int yupper = -minCount;
+
+        // cycle by neighbor rings
+        for (int i = minCount; i < maxRange.width; ++i)
+        {
+            // two cycles by y
+            int x = -(i + 1);
+            int y = yupper;
+
+            // left side
+            for (int j = -maxRange.height; j <= maxRange.height; ++j, ++ssCount, ++y)
+            {
+                ss[ssCount].x = x;
+                ss[ssCount].y = y;
+            }
+
+            y = yupper;
+            x = -x;
+
+            // right side
+            for (int j = -maxRange.height; j <= maxRange.height; ++j, ++ssCount, ++y)
+            {
+                ss[ssCount].x = x;
+                ss[ssCount].y = y;
+            }
+        }
+    }
+
+    const cudaStream_t stream = StreamAccessor::getStream(st);
+
+    ensureSizeIsEnough(1, ssCount, CV_16SC2, buf);
+    if (stream == 0)
+        cudaSafeCall( cudaMemcpy(buf.data, &ss[0], ssCount * sizeof(short2), cudaMemcpyHostToDevice) );
+    else
+        cudaSafeCall( cudaMemcpyAsync(buf.data, &ss[0], ssCount * sizeof(short2), cudaMemcpyHostToDevice, stream) );
+
+    const int maxX = prev.cols - blockSize.width;
+    const int maxY = prev.rows - blockSize.height;
+
+    const int SMALL_DIFF = 2;
+    const int BIG_DIFF = 128;
+
+    const int blSize = blockSize.area();
+    const int acceptLevel = blSize * SMALL_DIFF;
+    const int escapeLevel = blSize * BIG_DIFF;
+
+    optflowbm::calc(prev, curr, velx, vely,
+                    make_int2(blockSize.width, blockSize.height), make_int2(shiftSize.width, shiftSize.height), usePrevious,
+                    maxX, maxY, acceptLevel, escapeLevel, buf.ptr<short2>(), ssCount, stream);
+}
+
+namespace optflowbm_fast
+{
+    void get_buffer_size(int src_cols, int src_rows, int search_window, int block_window, int& buffer_cols, int& buffer_rows);
+
+    template <typename T>
+    void calc(PtrStepSzb I0, PtrStepSzb I1, PtrStepSzf velx, PtrStepSzf vely, PtrStepi buffer, int search_window, int block_window, cudaStream_t stream);
+}
+
+void cv::gpu::FastOpticalFlowBM::operator ()(const GpuMat& I0, const GpuMat& I1, GpuMat& flowx, GpuMat& flowy, int search_window, int block_window, Stream& stream)
+{
+    CV_Assert( I0.type() == CV_8UC1 );
+    CV_Assert( I1.size() == I0.size() && I1.type() == I0.type() );
+
+    int border_size = search_window / 2 + block_window / 2;
+    Size esize = I0.size() + Size(border_size, border_size) * 2;
+
+    ensureSizeIsEnough(esize, I0.type(), extended_I0);
+    ensureSizeIsEnough(esize, I0.type(), extended_I1);
+
+    copyMakeBorder(I0, extended_I0, border_size, border_size, border_size, border_size, cv::BORDER_DEFAULT, Scalar(), stream);
+    copyMakeBorder(I1, extended_I1, border_size, border_size, border_size, border_size, cv::BORDER_DEFAULT, Scalar(), stream);
+
+    GpuMat I0_hdr = extended_I0(Rect(Point2i(border_size, border_size), I0.size()));
+    GpuMat I1_hdr = extended_I1(Rect(Point2i(border_size, border_size), I0.size()));
+
+    int bcols, brows;
+    optflowbm_fast::get_buffer_size(I0.cols, I0.rows, search_window, block_window, bcols, brows);
+
+    ensureSizeIsEnough(brows, bcols, CV_32SC1, buffer);
+
+    flowx.create(I0.size(), CV_32FC1);
+    flowy.create(I0.size(), CV_32FC1);
+
+    optflowbm_fast::calc<uchar>(I0_hdr, I1_hdr, flowx, flowy, buffer, search_window, block_window, StreamAccessor::getStream(stream));
+}
+
+#endif // HAVE_CUDA

modules/gpu/test/test_optflow.cpp

@@ -445,4 +445,179 @@ INSTANTIATE_TEST_CASE_P(GPU_Video, OpticalFlowDual_TVL1, testing::Combine(
     ALL_DEVICES,
     WHOLE_SUBMAT));
 
+//////////////////////////////////////////////////////
+// OpticalFlowBM
+
+namespace
+{
+    void calcOpticalFlowBM(const cv::Mat& prev, const cv::Mat& curr,
+                           cv::Size bSize, cv::Size shiftSize, cv::Size maxRange, int usePrevious,
+                           cv::Mat& velx, cv::Mat& vely)
+    {
+        cv::Size sz((curr.cols - bSize.width + shiftSize.width)/shiftSize.width, (curr.rows - bSize.height + shiftSize.height)/shiftSize.height);
+
+        velx.create(sz, CV_32FC1);
+        vely.create(sz, CV_32FC1);
+
+        CvMat cvprev = prev;
+        CvMat cvcurr = curr;
+
+        CvMat cvvelx = velx;
+        CvMat cvvely = vely;
+
+        cvCalcOpticalFlowBM(&cvprev, &cvcurr, bSize, shiftSize, maxRange, usePrevious, &cvvelx, &cvvely);
+    }
+}
+
+struct OpticalFlowBM : testing::TestWithParam<cv::gpu::DeviceInfo>
+{
+};
+
+GPU_TEST_P(OpticalFlowBM, Accuracy)
+{
+    cv::gpu::DeviceInfo devInfo = GetParam();
+    cv::gpu::setDevice(devInfo.deviceID());
+
+    cv::Mat frame0 = readImage("opticalflow/rubberwhale1.png", cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(frame0.empty());
+
+    cv::Mat frame1 = readImage("opticalflow/rubberwhale2.png", cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(frame1.empty());
+
+    cv::Size block_size(16, 16);
+    cv::Size shift_size(1, 1);
+    cv::Size max_range(16, 16);
+
+    cv::gpu::GpuMat d_velx, d_vely, buf;
+    cv::gpu::calcOpticalFlowBM(loadMat(frame0), loadMat(frame1),
+                               block_size, shift_size, max_range, false,
+                               d_velx, d_vely, buf);
+
+    cv::Mat velx, vely;
+    calcOpticalFlowBM(frame0, frame1, block_size, shift_size, max_range, false, velx, vely);
+
+    EXPECT_MAT_NEAR(velx, d_velx, 0);
+    EXPECT_MAT_NEAR(vely, d_vely, 0);
+}
+
+INSTANTIATE_TEST_CASE_P(GPU_Video, OpticalFlowBM, ALL_DEVICES);
+
+//////////////////////////////////////////////////////
+// FastOpticalFlowBM
+
+namespace
+{
+    void FastOpticalFlowBM_gold(const cv::Mat_<uchar>& I0, const cv::Mat_<uchar>& I1, cv::Mat_<float>& velx, cv::Mat_<float>& vely, int search_window, int block_window)
+    {
+        velx.create(I0.size());
+        vely.create(I0.size());
+
+        int search_radius = search_window / 2;
+        int block_radius = block_window / 2;
+
+        for (int y = 0; y < I0.rows; ++y)
+        {
+            for (int x = 0; x < I0.cols; ++x)
+            {
+                int bestDist = std::numeric_limits<int>::max();
+                int bestDx = 0;
+                int bestDy = 0;
+
+                for (int dy = -search_radius; dy <= search_radius; ++dy)
+                {
+                    for (int dx = -search_radius; dx <= search_radius; ++dx)
+                    {
+                        int dist = 0;
+
+                        for (int by = -block_radius; by <= block_radius; ++by)
+                        {
+                            for (int bx = -block_radius; bx <= block_radius; ++bx)
+                            {
+                                int I0_val = I0(cv::borderInterpolate(y + by, I0.rows, cv::BORDER_DEFAULT), cv::borderInterpolate(x + bx, I0.cols, cv::BORDER_DEFAULT));
+                                int I1_val = I1(cv::borderInterpolate(y + dy + by, I0.rows, cv::BORDER_DEFAULT), cv::borderInterpolate(x + dx + bx, I0.cols, cv::BORDER_DEFAULT));
+
+                                dist += std::abs(I0_val - I1_val);
+                            }
+                        }
+
+                        if (dist < bestDist)
+                        {
+                            bestDist = dist;
+                            bestDx = dx;
+                            bestDy = dy;
+                        }
+                    }
+                }
+
+                velx(y, x) = (float) bestDx;
+                vely(y, x) = (float) bestDy;
+            }
+        }
+    }
+
+    double calc_rmse(const cv::Mat_<float>& flow1, const cv::Mat_<float>& flow2)
+    {
+        double sum = 0.0;
+
+        for (int y = 0; y < flow1.rows; ++y)
+        {
+            for (int x = 0; x < flow1.cols; ++x)
+            {
+                double diff = flow1(y, x) - flow2(y, x);
+                sum += diff * diff;
+            }
+        }
+
+        return std::sqrt(sum / flow1.size().area());
+    }
+}
+
+struct FastOpticalFlowBM : testing::TestWithParam<cv::gpu::DeviceInfo>
+{
+};
+
+GPU_TEST_P(FastOpticalFlowBM, Accuracy)
+{
+    const double MAX_RMSE = 0.6;
+
+    int search_window = 15;
+    int block_window = 5;
+
+    cv::gpu::DeviceInfo devInfo = GetParam();
+    cv::gpu::setDevice(devInfo.deviceID());
+
+    cv::Mat frame0 = readImage("opticalflow/rubberwhale1.png", cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(frame0.empty());
+
+    cv::Mat frame1 = readImage("opticalflow/rubberwhale2.png", cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(frame1.empty());
+
+    cv::Size smallSize(320, 240);
+    cv::Mat frame0_small;
+    cv::Mat frame1_small;
+
+    cv::resize(frame0, frame0_small, smallSize);
+    cv::resize(frame1, frame1_small, smallSize);
+
+    cv::gpu::GpuMat d_flowx;
+    cv::gpu::GpuMat d_flowy;
+    cv::gpu::FastOpticalFlowBM fastBM;
+
+    fastBM(loadMat(frame0_small), loadMat(frame1_small), d_flowx, d_flowy, search_window, block_window);
+
+    cv::Mat_<float> flowx;
+    cv::Mat_<float> flowy;
+    FastOpticalFlowBM_gold(frame0_small, frame1_small, flowx, flowy, search_window, block_window);
+
+    double err;
+
+    err = calc_rmse(flowx, cv::Mat(d_flowx));
+    EXPECT_LE(err, MAX_RMSE);
+
+    err = calc_rmse(flowy, cv::Mat(d_flowy));
+    EXPECT_LE(err, MAX_RMSE);
+}
+
+INSTANTIATE_TEST_CASE_P(GPU_Video, FastOpticalFlowBM, ALL_DEVICES);
+
 #endif // HAVE_CUDA