Merge pull request #14676 from 103yiran:103yiran-clang

std::rand() has no thread-safe guarantee.

backport commit: 147f3ebf0a

modules/androidcamera/src/camera_activity.cpp

@@ -5,8 +5,10 @@
 #include <android/log.h>
 #include <cctype>
 #include <string>
+#include <cstring>
 #include <vector>
 #include <algorithm>
+#include <functional>
 #include <opencv2/core/version.hpp>
 #include "camera_activity.hpp"
 #include "camera_wrapper.h"

modules/core/include/opencv2/core/version.hpp

@@ -50,7 +50,7 @@
 #define CV_VERSION_EPOCH    2
 #define CV_VERSION_MAJOR    4
 #define CV_VERSION_MINOR    13
-#define CV_VERSION_REVISION 6
+#define CV_VERSION_REVISION 7
 
 #define CVAUX_STR_EXP(__A)  #__A
 #define CVAUX_STR(__A)      CVAUX_STR_EXP(__A)

modules/core/test/test_eigen.cpp

@@ -164,7 +164,7 @@ void Core_EigenTest_32::run(int) { check_full(CV_32FC1); }
 void Core_EigenTest_64::run(int) { check_full(CV_64FC1); }
 
 Core_EigenTest::Core_EigenTest()
-: eps_val_32(1e-3f), eps_vec_32(1e-2f),
+: eps_val_32(1e-3f), eps_vec_32(2e-2f),
   eps_val_64(1e-4f), eps_vec_64(1e-3f), ntests(100) {}
 Core_EigenTest::~Core_EigenTest() {}
 

modules/flann/include/opencv2/flann/kdtree_index.h

@@ -132,7 +132,12 @@ public:
         /* Construct the randomized trees. */
         for (int i = 0; i < trees_; i++) {
             /* Randomize the order of vectors to allow for unbiased sampling. */
+#ifndef OPENCV_FLANN_USE_STD_RAND
+            cv::randShuffle(vind_);
+#else
             std::random_shuffle(vind_.begin(), vind_.end());
+#endif
+
             tree_roots_[i] = divideTree(&vind_[0], int(size_) );
         }
     }

modules/flann/include/opencv2/flann/lsh_index.h

@@ -136,7 +136,12 @@ public:
               indices.resize( feature_size_ * CHAR_BIT );
               for (size_t j = 0; j < feature_size_ * CHAR_BIT; ++j)
                   indices[j] = j;
+
+#ifndef OPENCV_FLANN_USE_STD_RAND
+              cv::randShuffle(indices);
+#else
               std::random_shuffle(indices.begin(), indices.end());
+#endif
             }
 
             lsh::LshTable<ElementType>& table = tables_[i];

modules/flann/include/opencv2/flann/random.h

@@ -40,13 +40,31 @@
 namespace cvflann
 {
 
+inline int rand()
+{
+#ifndef OPENCV_FLANN_USE_STD_RAND
+#   if INT_MAX == RAND_MAX
+    int v = cv::theRNG().next() & INT_MAX;
+#   else
+    int v = cv::theRNG().uniform(0, RAND_MAX + 1);
+#   endif
+#else
+    int v = std::rand();
+#endif // OPENCV_FLANN_USE_STD_RAND
+    return v;
+}
+
 /**
  * Seeds the random number generator
  *  @param seed Random seed
  */
 inline void seed_random(unsigned int seed)
 {
-    srand(seed);
+#ifndef OPENCV_FLANN_USE_STD_RAND
+    cv::theRNG() = cv::RNG(seed);
+#else
+    std::srand(seed);
+#endif
 }
 
 /*
@@ -60,7 +78,7 @@ inline void seed_random(unsigned int seed)
  */
 inline double rand_double(double high = 1.0, double low = 0)
 {
-    return low + ((high-low) * (std::rand() / (RAND_MAX + 1.0)));
+    return low + ((high-low) * (rand() / (RAND_MAX + 1.0)));
 }
 
 /**
@@ -71,7 +89,7 @@ inline double rand_double(double high = 1.0, double low = 0)
  */
 inline int rand_int(int high = RAND_MAX, int low = 0)
 {
-    return low + (int) ( double(high-low) * (std::rand() / (RAND_MAX + 1.0)));
+    return low + (int) ( double(high-low) * (rand() / (RAND_MAX + 1.0)));
 }
 
 /**
@@ -107,7 +125,11 @@ public:
         for (int i = 0; i < size_; ++i) vals_[i] = i;
 
         // shuffle the elements in the array
+#ifndef OPENCV_FLANN_USE_STD_RAND
+        cv::randShuffle(vals_);
+#else
         std::random_shuffle(vals_.begin(), vals_.end());
+#endif
 
         counter_ = 0;
     }

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

@@ -1824,6 +1824,9 @@ public:
     void sparse(const GpuMat& prevImg, const GpuMat& nextImg, const GpuMat& prevPts, GpuMat& nextPts,
         GpuMat& status, GpuMat* err = 0);
 
+    void sparse_multi(const GpuMat& prevImg, const GpuMat& nextImg, const GpuMat& prevPts, GpuMat& nextPts,
+        GpuMat& status, Stream& stream, GpuMat* err = 0);
+
     void dense(const GpuMat& prevImg, const GpuMat& nextImg, GpuMat& u, GpuMat& v, GpuMat* err = 0);
 
     void releaseMemory();

modules/gpu/perf/perf_video.cpp

@@ -303,6 +303,88 @@ PERF_TEST_P(ImagePair_Gray_NPts_WinSz_Levels_Iters, Video_PyrLKOpticalFlowSparse
     }
 }
 
+//////////////////////////////////////////////////////
+// PyrLKOpticalFlowSparseMulti
+
+#if defined(HAVE_TBB) && defined(HAVE_CUDA)
+
+DEF_PARAM_TEST(ImagePair_Gray_NPts_WinSz_Levels_Iters, pair_string, bool, int, int, int, int);
+
+PERF_TEST_P(ImagePair_Gray_NPts_WinSz_Levels_Iters, Video_PyrLKOpticalFlowSparseMulti,
+            Combine(Values<pair_string>(make_pair("gpu/opticalflow/frame0.png", "gpu/opticalflow/frame1.png")),
+                    Bool(),
+                    Values(8000),
+                    Values(21),
+                    Values(1, 3),
+                    Values(1, 30)))
+{
+    declare.time(20.0);
+
+    const pair_string imagePair = GET_PARAM(0);
+    const bool useGray = GET_PARAM(1);
+    const int points = GET_PARAM(2);
+    const int winSize = GET_PARAM(3);
+    const int levels = GET_PARAM(4);
+    const int iters = GET_PARAM(5);
+
+    const cv::Mat frame0 = readImage(imagePair.first, useGray ? cv::IMREAD_GRAYSCALE : cv::IMREAD_COLOR);
+    ASSERT_FALSE(frame0.empty());
+
+    const cv::Mat frame1 = readImage(imagePair.second, useGray ? cv::IMREAD_GRAYSCALE : cv::IMREAD_COLOR);
+    ASSERT_FALSE(frame1.empty());
+
+    cv::Mat gray_frame;
+    if (useGray)
+        gray_frame = frame0;
+    else
+        cv::cvtColor(frame0, gray_frame, cv::COLOR_BGR2GRAY);
+
+    cv::Mat pts;
+    cv::goodFeaturesToTrack(gray_frame, pts, points, 0.01, 0.0);
+
+    if (PERF_RUN_GPU())
+    {
+        const cv::gpu::GpuMat d_pts(pts.reshape(2, 1));
+
+        cv::gpu::PyrLKOpticalFlow d_pyrLK;
+        d_pyrLK.winSize = cv::Size(winSize, winSize);
+        d_pyrLK.maxLevel = levels - 1;
+        d_pyrLK.iters = iters;
+
+        const cv::gpu::GpuMat d_frame0(frame0);
+        const cv::gpu::GpuMat d_frame1(frame1);
+        cv::gpu::GpuMat nextPts;
+        cv::gpu::GpuMat status;
+
+        cv::gpu::Stream stream;
+        TEST_CYCLE()
+        {
+            d_pyrLK.sparse_multi(d_frame0, d_frame1, d_pts, nextPts, status, stream);
+            stream.waitForCompletion();
+        }
+
+        GPU_SANITY_CHECK(nextPts);
+        GPU_SANITY_CHECK(status);
+    }
+    else
+    {
+        cv::Mat nextPts;
+        cv::Mat status;
+
+        TEST_CYCLE()
+        {
+            cv::calcOpticalFlowPyrLK(frame0, frame1, pts, nextPts, status, cv::noArray(),
+                                     cv::Size(winSize, winSize), levels - 1,
+                                     cv::TermCriteria(cv::TermCriteria::COUNT + cv::TermCriteria::EPS, iters, 0.01));
+        }
+
+        CPU_SANITY_CHECK(nextPts);
+        CPU_SANITY_CHECK(status);
+    }
+}
+
+#endif // HAVE_TBB
+
 //////////////////////////////////////////////////////
 // PyrLKOpticalFlowDense
 

modules/gpu/src/cuda/pyrlk.cu

@@ -49,6 +49,10 @@
 #include "opencv2/gpu/device/vec_math.hpp"
 #include "opencv2/gpu/device/reduce.hpp"
 
+#include "opencv2/core/core.hpp"
+
+#include "cvconfig.h"
+
 using namespace cv::gpu;
 using namespace cv::gpu::device;
 
@@ -60,12 +64,54 @@ namespace pyrlk
     __constant__ int c_halfWin_y;
     __constant__ int c_iters;
 
-    texture<float, cudaTextureType2D, cudaReadModeElementType> tex_If(false, cudaFilterModeLinear, cudaAddressModeClamp);
+#define CUDA_CONSTANTS(index) \
-    texture<float4, cudaTextureType2D, cudaReadModeElementType> tex_If4(false, cudaFilterModeLinear, cudaAddressModeClamp);
+    __constant__ int c_winSize_x##index; \
-    texture<uchar, cudaTextureType2D, cudaReadModeElementType> tex_Ib(false, cudaFilterModePoint, cudaAddressModeClamp);
+    __constant__ int c_winSize_y##index; \
+    __constant__ int c_halfWin_x##index; \
+    __constant__ int c_halfWin_y##index; \
+    __constant__ int c_iters##index;
 
-    texture<float, cudaTextureType2D, cudaReadModeElementType> tex_Jf(false, cudaFilterModeLinear, cudaAddressModeClamp);
+    CUDA_CONSTANTS(0)
-    texture<float4, cudaTextureType2D, cudaReadModeElementType> tex_Jf4(false, cudaFilterModeLinear, cudaAddressModeClamp);
+    CUDA_CONSTANTS(1)
+    CUDA_CONSTANTS(2)
+    CUDA_CONSTANTS(3)
+    CUDA_CONSTANTS(4)
+
+    template <int index> struct c_multi_winSize_x;
+    template <int index> struct c_multi_winSize_y;
+    template <int index> struct c_multi_halfWin_x;
+    template <int index> struct c_multi_halfWin_y;
+    template <int index> struct c_multi_iters;
+
+#define CUDA_CONSTANTS_ACCESSOR(index) \
+    template <> struct c_multi_winSize_x<index> \
+    { static __device__ __forceinline__ int get(void){ return c_winSize_x##index;} }; \
+    template <> struct c_multi_winSize_y<index> \
+    { static __device__ __forceinline__ int get(void){ return c_winSize_y##index;} }; \
+    template <> struct c_multi_halfWin_x<index> \
+    { static __device__ __forceinline__ int get(void){ return c_halfWin_x##index;} }; \
+    template <> struct c_multi_halfWin_y<index> \
+    { static __device__ __forceinline__ int get(void){ return c_halfWin_y##index;} }; \
+    template <> struct c_multi_iters<index> \
+    { static __device__ __forceinline__ int get(void){ return c_iters##index;} };
+
+    CUDA_CONSTANTS_ACCESSOR(0)
+    CUDA_CONSTANTS_ACCESSOR(1)
+    CUDA_CONSTANTS_ACCESSOR(2)
+    CUDA_CONSTANTS_ACCESSOR(3)
+    CUDA_CONSTANTS_ACCESSOR(4)
+
+    texture<float, cudaTextureType2D, cudaReadModeElementType>
+            tex_If(false, cudaFilterModeLinear, cudaAddressModeClamp);
+    texture<float4, cudaTextureType2D, cudaReadModeElementType>
+            tex_If4(false, cudaFilterModeLinear, cudaAddressModeClamp);
+    texture<uchar, cudaTextureType2D, cudaReadModeElementType>
+            tex_Ib(false, cudaFilterModePoint, cudaAddressModeClamp);
+
+    texture<float, cudaTextureType2D, cudaReadModeElementType>
+            tex_Jf(false, cudaFilterModeLinear, cudaAddressModeClamp);
+    texture<float4, cudaTextureType2D, cudaReadModeElementType>
+            tex_Jf4(false, cudaFilterModeLinear, cudaAddressModeClamp);
 
     template <int cn> struct Tex_I;
     template <> struct Tex_I<1>
@@ -99,6 +145,57 @@ namespace pyrlk
         }
     };
 
+    //--------------------------------------------------------------------------
+
+#define CUDA_DECL_TEX_MULTI(texname, type, filtermode) \
+    texture<type, cudaTextureType2D, cudaReadModeElementType> \
+            texname##_multi0(false, filtermode, cudaAddressModeClamp); \
+    texture<type, cudaTextureType2D, cudaReadModeElementType> \
+            texname##_multi1(false, filtermode, cudaAddressModeClamp); \
+    texture<type, cudaTextureType2D, cudaReadModeElementType> \
+            texname##_multi2(false, filtermode, cudaAddressModeClamp); \
+    texture<type, cudaTextureType2D, cudaReadModeElementType> \
+            texname##_multi3(false, filtermode, cudaAddressModeClamp); \
+    texture<type, cudaTextureType2D, cudaReadModeElementType> \
+            texname##_multi4(false, filtermode, cudaAddressModeClamp); \
+
+    CUDA_DECL_TEX_MULTI(tex_If1, float, cudaFilterModeLinear)
+    CUDA_DECL_TEX_MULTI(tex_If4, float4, cudaFilterModeLinear)
+    CUDA_DECL_TEX_MULTI(tex_Ib1, uchar, cudaFilterModePoint)
+    CUDA_DECL_TEX_MULTI(tex_Jf1, float, cudaFilterModeLinear)
+    CUDA_DECL_TEX_MULTI(tex_Jf4, float4, cudaFilterModeLinear)
+
+    template <int cn, int index> struct Tex_I_multi;
+    template <int cn, int index> struct Tex_J_multi;
+    template <int cn, int index> struct Tex_B_multi;
+
+#define CUDA_DECL_TEX_MULTI_ACCESS(accessorname, texname, cn, returntype) \
+    template <> struct accessorname##_multi<cn, 0> \
+    { static __device__ __forceinline__ returntype read(float x, float y) \
+        { return tex2D(texname##cn##_multi0, x, y); } }; \
+    template <> struct accessorname##_multi<cn, 1> \
+    { static __device__ __forceinline__ returntype read(float x, float y) \
+        { return tex2D(texname##cn##_multi1, x, y); } }; \
+    template <> struct accessorname##_multi<cn, 2> \
+    { static __device__ __forceinline__ returntype read(float x, float y) \
+        { return tex2D(texname##cn##_multi2, x, y); } }; \
+    template <> struct accessorname##_multi<cn, 3> \
+    { static __device__ __forceinline__ returntype read(float x, float y) \
+        { return tex2D(texname##cn##_multi3, x, y); } }; \
+    template <> struct accessorname##_multi<cn, 4> \
+    { static __device__ __forceinline__ returntype read(float x, float y) \
+        { return tex2D(texname##cn##_multi4, x, y); } };
+
+    CUDA_DECL_TEX_MULTI_ACCESS(Tex_I, tex_If, 1, float)
+    CUDA_DECL_TEX_MULTI_ACCESS(Tex_I, tex_If, 4, float4)
+
+    CUDA_DECL_TEX_MULTI_ACCESS(Tex_B, tex_Ib, 1, uchar)
+
+    CUDA_DECL_TEX_MULTI_ACCESS(Tex_J, tex_Jf, 1, float)
+    CUDA_DECL_TEX_MULTI_ACCESS(Tex_J, tex_Jf, 4, float4)
+
+    //--------------------------------------------------------------------------
+
     __device__ __forceinline__ void accum(float& dst, float val)
     {
         dst += val;
@@ -309,6 +406,200 @@ namespace pyrlk
         }
     }
 
+#if defined(HAVE_TBB)
+    template <int cn, int index, int PATCH_X, int PATCH_Y, bool calcErr>
+    __global__ void sparseKernel_multi(const float2* prevPts, float2* nextPts, uchar* status, float* err, const int level, const int rows, const int cols)
+    {
+    #if __CUDA_ARCH__ <= 110
+        const int BLOCK_SIZE = 128;
+    #else
+        const int BLOCK_SIZE = 256;
+    #endif
+
+        __shared__ float smem1[BLOCK_SIZE];
+        __shared__ float smem2[BLOCK_SIZE];
+        __shared__ float smem3[BLOCK_SIZE];
+
+        const unsigned int tid = threadIdx.y * blockDim.x + threadIdx.x;
+
+        float2 prevPt = prevPts[blockIdx.x];
+        prevPt.x *= (1.0f / (1 << level));
+        prevPt.y *= (1.0f / (1 << level));
+
+        if (prevPt.x < 0 || prevPt.x >= cols || prevPt.y < 0 || prevPt.y >= rows)
+        {
+            if (tid == 0 && level == 0)
+                status[blockIdx.x] = 0;
+
+            return;
+        }
+
+        prevPt.x -= c_multi_halfWin_x<index>::get();
+        prevPt.y -= c_multi_halfWin_y<index>::get();
+
+        // extract the patch from the first image, compute covariation matrix of derivatives
+
+        float A11 = 0;
+        float A12 = 0;
+        float A22 = 0;
+
+        typedef typename TypeVec<float, cn>::vec_type work_type;
+
+        work_type I_patch   [PATCH_Y][PATCH_X];
+        work_type dIdx_patch[PATCH_Y][PATCH_X];
+        work_type dIdy_patch[PATCH_Y][PATCH_X];
+
+        for (int yBase = threadIdx.y, i = 0; yBase < c_multi_winSize_y<index>::get(); yBase += blockDim.y, ++i)
+        {
+            for (int xBase = threadIdx.x, j = 0; xBase < c_multi_winSize_x<index>::get(); xBase += blockDim.x, ++j)
+            {
+                float x = prevPt.x + xBase + 0.5f;
+                float y = prevPt.y + yBase + 0.5f;
+
+                I_patch[i][j] = Tex_I_multi<cn,index>::read(x, y);
+
+                // Sharr Deriv
+
+                work_type dIdx = 3.0f * Tex_I_multi<cn,index>::read(x+1, y-1) + 10.0f * Tex_I_multi<cn,index>::read(x+1, y) + 3.0f * Tex_I_multi<cn,index>::read(x+1, y+1) -
+                                 (3.0f * Tex_I_multi<cn,index>::read(x-1, y-1) + 10.0f * Tex_I_multi<cn,index>::read(x-1, y) + 3.0f * Tex_I_multi<cn,index>::read(x-1, y+1));
+
+                work_type dIdy = 3.0f * Tex_I_multi<cn,index>::read(x-1, y+1) + 10.0f * Tex_I_multi<cn,index>::read(x, y+1) + 3.0f * Tex_I_multi<cn,index>::read(x+1, y+1) -
+                                (3.0f * Tex_I_multi<cn,index>::read(x-1, y-1) + 10.0f * Tex_I_multi<cn,index>::read(x, y-1) + 3.0f * Tex_I_multi<cn,index>::read(x+1, y-1));
+
+                dIdx_patch[i][j] = dIdx;
+                dIdy_patch[i][j] = dIdy;
+
+                accum(A11, dIdx * dIdx);
+                accum(A12, dIdx * dIdy);
+                accum(A22, dIdy * dIdy);
+            }
+        }
+
+        reduce<BLOCK_SIZE>(smem_tuple(smem1, smem2, smem3), thrust::tie(A11, A12, A22), tid, thrust::make_tuple(plus<float>(), plus<float>(), plus<float>()));
+
+    #if __CUDA_ARCH__ >= 300
+        if (tid == 0)
+        {
+            smem1[0] = A11;
+            smem2[0] = A12;
+            smem3[0] = A22;
+        }
+    #endif
+
+        __syncthreads();
+
+        A11 = smem1[0];
+        A12 = smem2[0];
+        A22 = smem3[0];
+
+        float D = A11 * A22 - A12 * A12;
+
+        if (abs_(D) < numeric_limits<float>::epsilon())
+        {
+            if (tid == 0 && level == 0)
+                status[blockIdx.x] = 0;
+
+            return;
+        }
+
+        D = 1.f / D;
+
+        A11 *= D;
+        A12 *= D;
+        A22 *= D;
+
+        float2 nextPt = nextPts[blockIdx.x];
+        nextPt.x *= 2.f;
+        nextPt.y *= 2.f;
+
+        nextPt.x -= c_multi_halfWin_x<index>::get();
+        nextPt.y -= c_multi_halfWin_y<index>::get();
+
+        for (int k = 0; k < c_multi_iters<index>::get(); ++k)
+        {
+            if (nextPt.x < -c_multi_halfWin_x<index>::get() || nextPt.x >= cols || nextPt.y < -c_multi_halfWin_y<index>::get() || nextPt.y >= rows)
+            {
+                if (tid == 0 && level == 0)
+                    status[blockIdx.x] = 0;
+
+                return;
+            }
+
+            float b1 = 0;
+            float b2 = 0;
+
+            for (int y = threadIdx.y, i = 0; y < c_multi_winSize_y<index>::get(); y += blockDim.y, ++i)
+            {
+                for (int x = threadIdx.x, j = 0; x < c_multi_winSize_x<index>::get(); x += blockDim.x, ++j)
+                {
+                    work_type I_val = I_patch[i][j];
+                    work_type J_val = Tex_J_multi<cn,index>::read(nextPt.x + x + 0.5f, nextPt.y + y + 0.5f);
+
+                    work_type diff = (J_val - I_val) * 32.0f;
+
+                    accum(b1, diff * dIdx_patch[i][j]);
+                    accum(b2, diff * dIdy_patch[i][j]);
+                }
+            }
+
+            reduce<BLOCK_SIZE>(smem_tuple(smem1, smem2), thrust::tie(b1, b2), tid, thrust::make_tuple(plus<float>(), plus<float>()));
+
+        #if __CUDA_ARCH__ >= 300
+            if (tid == 0)
+            {
+                smem1[0] = b1;
+                smem2[0] = b2;
+            }
+        #endif
+
+            __syncthreads();
+
+            b1 = smem1[0];
+            b2 = smem2[0];
+
+            float2 delta;
+            delta.x = A12 * b2 - A22 * b1;
+            delta.y = A12 * b1 - A11 * b2;
+
+            nextPt.x += delta.x;
+            nextPt.y += delta.y;
+
+            if (::fabs(delta.x) < 0.01f && ::fabs(delta.y) < 0.01f)
+                break;
+        }
+
+        float errval = 0;
+        if (calcErr)
+        {
+            for (int y = threadIdx.y, i = 0; y < c_multi_winSize_y<index>::get(); y += blockDim.y, ++i)
+            {
+                for (int x = threadIdx.x, j = 0; x < c_multi_winSize_x<index>::get(); x += blockDim.x, ++j)
+                {
+                    work_type I_val = I_patch[i][j];
+                    work_type J_val = Tex_J_multi<cn,index>::read(nextPt.x + x + 0.5f, nextPt.y + y + 0.5f);
+
+                    work_type diff = J_val - I_val;
+
+                    accum(errval, abs_(diff));
+                }
+            }
+
+            reduce<BLOCK_SIZE>(smem1, errval, tid, plus<float>());
+        }
+
+        if (tid == 0)
+        {
+            nextPt.x += c_multi_halfWin_x<index>::get();
+            nextPt.y += c_multi_halfWin_y<index>::get();
+
+            nextPts[blockIdx.x] = nextPt;
+
+            if (calcErr)
+                err[blockIdx.x] = static_cast<float>(errval) / (cn * c_multi_winSize_x<index>::get() * c_multi_winSize_y<index>::get());
+        }
+    }
+#endif // defined(HAVE_TBB)
+
     template <int cn, int PATCH_X, int PATCH_Y>
     void sparse_caller(int rows, int cols, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
                        int level, dim3 block, cudaStream_t stream)
@@ -326,6 +617,26 @@ namespace pyrlk
             cudaSafeCall( cudaDeviceSynchronize() );
     }
 
+#if defined(HAVE_TBB)
+    template <int cn, int index, int PATCH_X, int PATCH_Y>
+    void sparse_caller_multi(int rows, int cols, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
+                       int level, dim3 block, cudaStream_t stream)
+    {
+        dim3 grid(ptcount);
+
+        if (level == 0 && err)
+            sparseKernel_multi<cn, index, PATCH_X, PATCH_Y, true><<<grid, block>>>(prevPts, nextPts, status, err, level, rows, cols);
+        else
+            sparseKernel_multi<cn, index, PATCH_X, PATCH_Y, false><<<grid, block>>>(prevPts, nextPts, status, err, level, rows, cols);
+
+        cudaSafeCall( cudaGetLastError() );
+
+        if (stream == 0)
+            cudaSafeCall( cudaDeviceSynchronize() );
+    }
+
+#endif // defined(HAVE_TBB)
+
     template <bool calcErr>
     __global__ void denseKernel(PtrStepf u, PtrStepf v, const PtrStepf prevU, const PtrStepf prevV, PtrStepf err, const int rows, const int cols)
     {
@@ -484,6 +795,30 @@ namespace pyrlk
         cudaSafeCall( cudaMemcpyToSymbol(c_iters, &iters, sizeof(int)) );
     }
 
+#if defined(HAVE_TBB)
+    void loadConstants_multi(int2 winSize, int iters, int index, cudaStream_t stream = 0)
+    {
+        int2 halfWin;
+#define COPY_TO_SYMBOL_CALL(index) \
+        cudaSafeCall( cudaMemcpyToSymbolAsync(c_winSize_x##index, &winSize.x, sizeof(int), 0, cudaMemcpyHostToDevice, stream) ); \
+        cudaSafeCall( cudaMemcpyToSymbolAsync(c_winSize_y##index, &winSize.y, sizeof(int), 0, cudaMemcpyHostToDevice, stream) ); \
+        halfWin = make_int2((winSize.x - 1) / 2, (winSize.y - 1) / 2); \
+        cudaSafeCall( cudaMemcpyToSymbolAsync(c_halfWin_x##index, &halfWin.x, sizeof(int), 0, cudaMemcpyHostToDevice, stream) ); \
+        cudaSafeCall( cudaMemcpyToSymbolAsync(c_halfWin_y##index, &halfWin.y, sizeof(int), 0, cudaMemcpyHostToDevice, stream) ); \
+        cudaSafeCall( cudaMemcpyToSymbolAsync(c_iters##index, &iters, sizeof(int), 0, cudaMemcpyHostToDevice, stream) );
+
+        switch(index)
+        {
+            case 0: COPY_TO_SYMBOL_CALL(0) break;
+            case 1: COPY_TO_SYMBOL_CALL(1) break;
+            case 2: COPY_TO_SYMBOL_CALL(2) break;
+            case 3: COPY_TO_SYMBOL_CALL(3) break;
+            case 4: COPY_TO_SYMBOL_CALL(4) break;
+            default: CV_Error(CV_StsBadArg, "invalid execution line index"); break;
+        }
+    }
+#endif // defined(HAVE_TBB)
+
     void sparse1(PtrStepSzf I, PtrStepSzf J, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
                  int level, dim3 block, dim3 patch, cudaStream_t stream)
     {
@@ -528,6 +863,161 @@ namespace pyrlk
             level, block, stream);
     }
 
+#if defined(HAVE_TBB)
+    void sparse1_multi(PtrStepSzf I, PtrStepSzf J, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
+                 int level, dim3 block, dim3 patch, cudaStream_t stream, int index)
+    {
+        typedef void (*func_t)(int rows, int cols, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
+                               int level, dim3 block, cudaStream_t stream);
+
+        static const func_t funcs[5][5][5] =
+        {
+            { // index 0
+                {sparse_caller_multi<1, 0, 1, 1>, sparse_caller_multi<1, 0, 2, 1>, sparse_caller_multi<1, 0, 3, 1>, sparse_caller_multi<1, 0, 4, 1>, sparse_caller_multi<1, 0, 5, 1>},
+                {sparse_caller_multi<1, 0, 1, 2>, sparse_caller_multi<1, 0, 2, 2>, sparse_caller_multi<1, 0, 3, 2>, sparse_caller_multi<1, 0, 4, 2>, sparse_caller_multi<1, 0, 5, 2>},
+                {sparse_caller_multi<1, 0, 1, 3>, sparse_caller_multi<1, 0, 2, 3>, sparse_caller_multi<1, 0, 3, 3>, sparse_caller_multi<1, 0, 4, 3>, sparse_caller_multi<1, 0, 5, 3>},
+                {sparse_caller_multi<1, 0, 1, 4>, sparse_caller_multi<1, 0, 2, 4>, sparse_caller_multi<1, 0, 3, 4>, sparse_caller_multi<1, 0, 4, 4>, sparse_caller_multi<1, 0, 5, 4>},
+                {sparse_caller_multi<1, 0, 1, 5>, sparse_caller_multi<1, 0, 2, 5>, sparse_caller_multi<1, 0, 3, 5>, sparse_caller_multi<1, 0, 4, 5>, sparse_caller_multi<1, 0, 5, 5>}
+            },
+            { // index 1
+                {sparse_caller_multi<1, 1, 1, 1>, sparse_caller_multi<1, 1, 2, 1>, sparse_caller_multi<1, 1, 3, 1>, sparse_caller_multi<1, 1, 4, 1>, sparse_caller_multi<1, 1, 5, 1>},
+                {sparse_caller_multi<1, 1, 1, 2>, sparse_caller_multi<1, 1, 2, 2>, sparse_caller_multi<1, 1, 3, 2>, sparse_caller_multi<1, 1, 4, 2>, sparse_caller_multi<1, 1, 5, 2>},
+                {sparse_caller_multi<1, 1, 1, 3>, sparse_caller_multi<1, 1, 2, 3>, sparse_caller_multi<1, 1, 3, 3>, sparse_caller_multi<1, 1, 4, 3>, sparse_caller_multi<1, 1, 5, 3>},
+                {sparse_caller_multi<1, 1, 1, 4>, sparse_caller_multi<1, 1, 2, 4>, sparse_caller_multi<1, 1, 3, 4>, sparse_caller_multi<1, 1, 4, 4>, sparse_caller_multi<1, 1, 5, 4>},
+                {sparse_caller_multi<1, 1, 1, 5>, sparse_caller_multi<1, 1, 2, 5>, sparse_caller_multi<1, 1, 3, 5>, sparse_caller_multi<1, 1, 4, 5>, sparse_caller_multi<1, 1, 5, 5>}
+            },
+            { // index 2
+                {sparse_caller_multi<1, 2, 1, 1>, sparse_caller_multi<1, 2, 2, 1>, sparse_caller_multi<1, 2, 3, 1>, sparse_caller_multi<1, 2, 4, 1>, sparse_caller_multi<1, 2, 5, 1>},
+                {sparse_caller_multi<1, 2, 1, 2>, sparse_caller_multi<1, 2, 2, 2>, sparse_caller_multi<1, 2, 3, 2>, sparse_caller_multi<1, 2, 4, 2>, sparse_caller_multi<1, 2, 5, 2>},
+                {sparse_caller_multi<1, 2, 1, 3>, sparse_caller_multi<1, 2, 2, 3>, sparse_caller_multi<1, 2, 3, 3>, sparse_caller_multi<1, 2, 4, 3>, sparse_caller_multi<1, 2, 5, 3>},
+                {sparse_caller_multi<1, 2, 1, 4>, sparse_caller_multi<1, 2, 2, 4>, sparse_caller_multi<1, 2, 3, 4>, sparse_caller_multi<1, 2, 4, 4>, sparse_caller_multi<1, 2, 5, 4>},
+                {sparse_caller_multi<1, 2, 1, 5>, sparse_caller_multi<1, 2, 2, 5>, sparse_caller_multi<1, 2, 3, 5>, sparse_caller_multi<1, 2, 4, 5>, sparse_caller_multi<1, 2, 5, 5>}
+            },
+            { // index 3
+                {sparse_caller_multi<1, 3, 1, 1>, sparse_caller_multi<1, 3, 2, 1>, sparse_caller_multi<1, 3, 3, 1>, sparse_caller_multi<1, 3, 4, 1>, sparse_caller_multi<1, 3, 5, 1>},
+                {sparse_caller_multi<1, 3, 1, 2>, sparse_caller_multi<1, 3, 2, 2>, sparse_caller_multi<1, 3, 3, 2>, sparse_caller_multi<1, 3, 4, 2>, sparse_caller_multi<1, 3, 5, 2>},
+                {sparse_caller_multi<1, 3, 1, 3>, sparse_caller_multi<1, 3, 2, 3>, sparse_caller_multi<1, 3, 3, 3>, sparse_caller_multi<1, 3, 4, 3>, sparse_caller_multi<1, 3, 5, 3>},
+                {sparse_caller_multi<1, 3, 1, 4>, sparse_caller_multi<1, 3, 2, 4>, sparse_caller_multi<1, 3, 3, 4>, sparse_caller_multi<1, 3, 4, 4>, sparse_caller_multi<1, 3, 5, 4>},
+                {sparse_caller_multi<1, 3, 1, 5>, sparse_caller_multi<1, 3, 2, 5>, sparse_caller_multi<1, 3, 3, 5>, sparse_caller_multi<1, 3, 4, 5>, sparse_caller_multi<1, 3, 5, 5>}
+            },
+            { // index 4
+                {sparse_caller_multi<1, 4, 1, 1>, sparse_caller_multi<1, 4, 2, 1>, sparse_caller_multi<1, 4, 3, 1>, sparse_caller_multi<1, 4, 4, 1>, sparse_caller_multi<1, 4, 5, 1>},
+                {sparse_caller_multi<1, 4, 1, 2>, sparse_caller_multi<1, 4, 2, 2>, sparse_caller_multi<1, 4, 3, 2>, sparse_caller_multi<1, 4, 4, 2>, sparse_caller_multi<1, 4, 5, 2>},
+                {sparse_caller_multi<1, 4, 1, 3>, sparse_caller_multi<1, 4, 2, 3>, sparse_caller_multi<1, 4, 3, 3>, sparse_caller_multi<1, 4, 4, 3>, sparse_caller_multi<1, 4, 5, 3>},
+                {sparse_caller_multi<1, 4, 1, 4>, sparse_caller_multi<1, 4, 2, 4>, sparse_caller_multi<1, 4, 3, 4>, sparse_caller_multi<1, 4, 4, 4>, sparse_caller_multi<1, 4, 5, 4>},
+                {sparse_caller_multi<1, 4, 1, 5>, sparse_caller_multi<1, 4, 2, 5>, sparse_caller_multi<1, 4, 3, 5>, sparse_caller_multi<1, 4, 4, 5>, sparse_caller_multi<1, 4, 5, 5>}
+            }
+        };
+
+        switch(index)
+        {
+            case 0:
+                bindTexture(&tex_If1_multi0, I);
+                bindTexture(&tex_Jf1_multi0, J);
+                break;
+            case 1:
+                bindTexture(&tex_If1_multi1, I);
+                bindTexture(&tex_Jf1_multi1, J);
+                break;
+            case 2:
+                bindTexture(&tex_If1_multi2, I);
+                bindTexture(&tex_Jf1_multi2, J);
+                break;
+            case 3:
+                bindTexture(&tex_If1_multi3, I);
+                bindTexture(&tex_Jf1_multi3, J);
+                break;
+            case 4:
+                bindTexture(&tex_If1_multi4, I);
+                bindTexture(&tex_Jf1_multi4, J);
+                break;
+            default:
+                CV_Error(CV_StsBadArg, "invalid execution line index");
+                break;
+        }
+
+        funcs[index][patch.y - 1][patch.x - 1](I.rows, I.cols, prevPts, nextPts, status, err, ptcount,
+            level, block, stream);
+    }
+
+    void sparse4_multi(PtrStepSz<float4> I, PtrStepSz<float4> J, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
+                 int level, dim3 block, dim3 patch, cudaStream_t stream, int index)
+    {
+        typedef void (*func_t)(int rows, int cols, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
+                               int level, dim3 block, cudaStream_t stream);
+
+        static const func_t funcs[5][5][5] =
+        {
+            { // index 0
+                {sparse_caller_multi<4, 0, 1, 1>, sparse_caller_multi<4, 0, 2, 1>, sparse_caller_multi<4, 0, 3, 1>, sparse_caller_multi<4, 0, 4, 1>, sparse_caller_multi<4, 0, 5, 1>},
+                {sparse_caller_multi<4, 0, 1, 2>, sparse_caller_multi<4, 0, 2, 2>, sparse_caller_multi<4, 0, 3, 2>, sparse_caller_multi<4, 0, 4, 2>, sparse_caller_multi<4, 0, 5, 2>},
+                {sparse_caller_multi<4, 0, 1, 3>, sparse_caller_multi<4, 0, 2, 3>, sparse_caller_multi<4, 0, 3, 3>, sparse_caller_multi<4, 0, 4, 3>, sparse_caller_multi<4, 0, 5, 3>},
+                {sparse_caller_multi<4, 0, 1, 4>, sparse_caller_multi<4, 0, 2, 4>, sparse_caller_multi<4, 0, 3, 4>, sparse_caller_multi<4, 0, 4, 4>, sparse_caller_multi<4, 0, 5, 4>},
+                {sparse_caller_multi<4, 0, 1, 5>, sparse_caller_multi<4, 0, 2, 5>, sparse_caller_multi<4, 0, 3, 5>, sparse_caller_multi<4, 0, 4, 5>, sparse_caller_multi<4, 0, 5, 5>}
+            },
+            { // index 1
+                {sparse_caller_multi<4, 1, 1, 1>, sparse_caller_multi<4, 1, 2, 1>, sparse_caller_multi<4, 1, 3, 1>, sparse_caller_multi<4, 1, 4, 1>, sparse_caller_multi<4, 1, 5, 1>},
+                {sparse_caller_multi<4, 1, 1, 2>, sparse_caller_multi<4, 1, 2, 2>, sparse_caller_multi<4, 1, 3, 2>, sparse_caller_multi<4, 1, 4, 2>, sparse_caller_multi<4, 1, 5, 2>},
+                {sparse_caller_multi<4, 1, 1, 3>, sparse_caller_multi<4, 1, 2, 3>, sparse_caller_multi<4, 1, 3, 3>, sparse_caller_multi<4, 1, 4, 3>, sparse_caller_multi<4, 1, 5, 3>},
+                {sparse_caller_multi<4, 1, 1, 4>, sparse_caller_multi<4, 1, 2, 4>, sparse_caller_multi<4, 1, 3, 4>, sparse_caller_multi<4, 1, 4, 4>, sparse_caller_multi<4, 1, 5, 4>},
+                {sparse_caller_multi<4, 1, 1, 5>, sparse_caller_multi<4, 1, 2, 5>, sparse_caller_multi<4, 1, 3, 5>, sparse_caller_multi<4, 1, 4, 5>, sparse_caller_multi<4, 1, 5, 5>}
+            },
+            { // index 2
+                {sparse_caller_multi<4, 2, 1, 1>, sparse_caller_multi<4, 2, 2, 1>, sparse_caller_multi<4, 2, 3, 1>, sparse_caller_multi<4, 2, 4, 1>, sparse_caller_multi<4, 2, 5, 1>},
+                {sparse_caller_multi<4, 2, 1, 2>, sparse_caller_multi<4, 2, 2, 2>, sparse_caller_multi<4, 2, 3, 2>, sparse_caller_multi<4, 2, 4, 2>, sparse_caller_multi<4, 2, 5, 2>},
+                {sparse_caller_multi<4, 2, 1, 3>, sparse_caller_multi<4, 2, 2, 3>, sparse_caller_multi<4, 2, 3, 3>, sparse_caller_multi<4, 2, 4, 3>, sparse_caller_multi<4, 2, 5, 3>},
+                {sparse_caller_multi<4, 2, 1, 4>, sparse_caller_multi<4, 2, 2, 4>, sparse_caller_multi<4, 2, 3, 4>, sparse_caller_multi<4, 2, 4, 4>, sparse_caller_multi<4, 2, 5, 4>},
+                {sparse_caller_multi<4, 2, 1, 5>, sparse_caller_multi<4, 2, 2, 5>, sparse_caller_multi<4, 2, 3, 5>, sparse_caller_multi<4, 2, 4, 5>, sparse_caller_multi<4, 2, 5, 5>}
+            },
+            { // index 3
+                {sparse_caller_multi<4, 3, 1, 1>, sparse_caller_multi<4, 3, 2, 1>, sparse_caller_multi<4, 3, 3, 1>, sparse_caller_multi<4, 3, 4, 1>, sparse_caller_multi<4, 3, 5, 1>},
+                {sparse_caller_multi<4, 3, 1, 2>, sparse_caller_multi<4, 3, 2, 2>, sparse_caller_multi<4, 3, 3, 2>, sparse_caller_multi<4, 3, 4, 2>, sparse_caller_multi<4, 3, 5, 2>},
+                {sparse_caller_multi<4, 3, 1, 3>, sparse_caller_multi<4, 3, 2, 3>, sparse_caller_multi<4, 3, 3, 3>, sparse_caller_multi<4, 3, 4, 3>, sparse_caller_multi<4, 3, 5, 3>},
+                {sparse_caller_multi<4, 3, 1, 4>, sparse_caller_multi<4, 3, 2, 4>, sparse_caller_multi<4, 3, 3, 4>, sparse_caller_multi<4, 3, 4, 4>, sparse_caller_multi<4, 3, 5, 4>},
+                {sparse_caller_multi<4, 3, 1, 5>, sparse_caller_multi<4, 3, 2, 5>, sparse_caller_multi<4, 3, 3, 5>, sparse_caller_multi<4, 3, 4, 5>, sparse_caller_multi<4, 3, 5, 5>}
+            },
+            { // index 4
+                {sparse_caller_multi<4, 4, 1, 1>, sparse_caller_multi<4, 4, 2, 1>, sparse_caller_multi<4, 4, 3, 1>, sparse_caller_multi<4, 4, 4, 1>, sparse_caller_multi<4, 4, 5, 1>},
+                {sparse_caller_multi<4, 4, 1, 2>, sparse_caller_multi<4, 4, 2, 2>, sparse_caller_multi<4, 4, 3, 2>, sparse_caller_multi<4, 4, 4, 2>, sparse_caller_multi<4, 4, 5, 2>},
+                {sparse_caller_multi<4, 4, 1, 3>, sparse_caller_multi<4, 4, 2, 3>, sparse_caller_multi<4, 4, 3, 3>, sparse_caller_multi<4, 4, 4, 3>, sparse_caller_multi<4, 4, 5, 3>},
+                {sparse_caller_multi<4, 4, 1, 4>, sparse_caller_multi<4, 4, 2, 4>, sparse_caller_multi<4, 4, 3, 4>, sparse_caller_multi<4, 4, 4, 4>, sparse_caller_multi<4, 4, 5, 4>},
+                {sparse_caller_multi<4, 4, 1, 5>, sparse_caller_multi<4, 4, 2, 5>, sparse_caller_multi<4, 4, 3, 5>, sparse_caller_multi<4, 4, 4, 5>, sparse_caller_multi<4, 4, 5, 5>}
+            }
+        };
+
+        switch(index)
+        {
+            case 0:
+                bindTexture(&tex_If4_multi0, I);
+                bindTexture(&tex_Jf4_multi0, J);
+                break;
+            case 1:
+                bindTexture(&tex_If4_multi1, I);
+                bindTexture(&tex_Jf4_multi1, J);
+                break;
+            case 2:
+                bindTexture(&tex_If4_multi2, I);
+                bindTexture(&tex_Jf4_multi2, J);
+                break;
+            case 3:
+                bindTexture(&tex_If4_multi3, I);
+                bindTexture(&tex_Jf4_multi3, J);
+                break;
+            case 4:
+                bindTexture(&tex_If4_multi4, I);
+                bindTexture(&tex_Jf4_multi4, J);
+                break;
+            default:
+                CV_Error(CV_StsBadArg, "invalid execution line index");
+                break;
+        }
+
+        funcs[index][patch.y - 1][patch.x - 1](I.rows, I.cols, prevPts, nextPts, status, err, ptcount,
+            level, block, stream);
+    }
+
+#endif // defined(HAVE_TBB)
+
     void dense(PtrStepSzb I, PtrStepSzf J, PtrStepSzf u, PtrStepSzf v, PtrStepSzf prevU, PtrStepSzf prevV, PtrStepSzf err, int2 winSize, cudaStream_t stream)
     {
         dim3 block(16, 16);

modules/gpu/src/pyrlk.cpp

@@ -42,12 +42,18 @@
 
 #include "precomp.hpp"
 
+#ifdef HAVE_TBB
+#include <tbb/compat/condition_variable>
+#include <tbb/mutex.h>
+#endif
+
 using namespace std;
 using namespace cv;
 using namespace cv::gpu;
 
 #if !defined (HAVE_CUDA) || defined (CUDA_DISABLER)
 
+
 cv::gpu::PyrLKOpticalFlow::PyrLKOpticalFlow() { throw_nogpu(); }
 void cv::gpu::PyrLKOpticalFlow::sparse(const GpuMat&, const GpuMat&, const GpuMat&, GpuMat&, GpuMat&, GpuMat*) { throw_nogpu(); }
 void cv::gpu::PyrLKOpticalFlow::dense(const GpuMat&, const GpuMat&, GpuMat&, GpuMat&, GpuMat*) { throw_nogpu(); }
@@ -64,6 +70,23 @@ namespace pyrlk
     void sparse4(PtrStepSz<float4> I, PtrStepSz<float4> J, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
                  int level, dim3 block, dim3 patch, cudaStream_t stream = 0);
 
+
+#if !defined(HAVE_TBB)
+#define throw_notbb() CV_Error(CV_StsNotImplemented, "The library is compiled without TBB support")
+    void loadConstants_multi(int2, int, int, cudaStream_t) { throw_notbb(); }
+    void sparse1_multi(PtrStepSzf, PtrStepSzf, const float2*, float2*, uchar*, float*, int,
+                 int, dim3, dim3, cudaStream_t, int) { throw_notbb(); }
+    void sparse4_multi(PtrStepSz<float4>, PtrStepSz<float4>, const float2*, float2*, uchar*, float*, int,
+                 int, dim3, dim3, cudaStream_t, int) { throw_notbb(); }
+#else
+    void loadConstants_multi(int2 winSize, int iters, int index = 0, cudaStream_t stream = 0);
+
+    void sparse1_multi(PtrStepSzf I, PtrStepSzf J, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
+                 int level, dim3 block, dim3 patch, cudaStream_t stream = 0, int index = 0);
+    void sparse4_multi(PtrStepSz<float4> I, PtrStepSz<float4> J, const float2* prevPts, float2* nextPts, uchar* status, float* err, int ptcount,
+                 int level, dim3 block, dim3 patch, cudaStream_t stream = 0, int index = 0);
+#endif
+
     void dense(PtrStepSzb I, PtrStepSzf J, PtrStepSzf u, PtrStepSzf v, PtrStepSzf prevU, PtrStepSzf prevV,
                PtrStepSzf err, int2 winSize, cudaStream_t stream = 0);
 }
@@ -98,7 +121,9 @@ namespace
     }
 }
 
-void cv::gpu::PyrLKOpticalFlow::sparse(const GpuMat& prevImg, const GpuMat& nextImg, const GpuMat& prevPts, GpuMat& nextPts, GpuMat& status, GpuMat* err)
+void cv::gpu::PyrLKOpticalFlow::sparse(const GpuMat& prevImg,
+                        const GpuMat& nextImg, const GpuMat& prevPts,
+                        GpuMat& nextPts, GpuMat& status, GpuMat* err)
 {
     if (prevPts.empty())
     {
@@ -181,6 +206,137 @@ void cv::gpu::PyrLKOpticalFlow::sparse(const GpuMat& prevImg, const GpuMat& next
     }
 }
 
+#ifdef HAVE_TBB
+//--------------------------------------------------------------------------
+// Multi-threading support
+
+static bool index_vector_use[5] = {true, true, true, true, true}; // all free
+static tbb::mutex s_PyrLKOpticalFlow_Mutex;
+static condition_variable s_PyrLKOpticalFlow_ConditionVariable;
+
+void cv::gpu::PyrLKOpticalFlow::sparse_multi(const GpuMat& prevImg,
+        const GpuMat& nextImg, const GpuMat& prevPts, GpuMat& nextPts,
+        GpuMat& status, Stream& stream, GpuMat* err)
+{
+    if (prevPts.empty())
+    {
+        nextPts.release();
+        status.release();
+        if (err) err->release();
+        return;
+    }
+
+    dim3 block, patch;
+    calcPatchSize(winSize, block, patch);
+
+    CV_Assert(prevImg.channels() == 1 || prevImg.channels() == 3 || prevImg.channels() == 4);
+    CV_Assert(prevImg.size() == nextImg.size() && prevImg.type() == nextImg.type());
+    CV_Assert(maxLevel >= 0);
+    CV_Assert(winSize.width > 2 && winSize.height > 2);
+    CV_Assert(patch.x > 0 && patch.x < 6 && patch.y > 0 && patch.y < 6);
+    CV_Assert(prevPts.rows == 1 && prevPts.type() == CV_32FC2);
+
+    if (useInitialFlow)
+        CV_Assert(nextPts.size() == prevPts.size() && nextPts.type() == CV_32FC2);
+    else
+        ensureSizeIsEnough(1, prevPts.cols, prevPts.type(), nextPts);
+
+    GpuMat temp1 = (useInitialFlow ? nextPts : prevPts).reshape(1);
+    GpuMat temp2 = nextPts.reshape(1);
+    multiply(temp1, Scalar::all(1.0 / (1 << maxLevel) / 2.0), temp2);
+
+    ensureSizeIsEnough(1, prevPts.cols, CV_8UC1, status);
+    status.setTo(Scalar::all(1));
+
+    if (err)
+        ensureSizeIsEnough(1, prevPts.cols, CV_32FC1, *err);
+
+    // build the image pyramids.
+
+    prevPyr_.resize(maxLevel + 1);
+    nextPyr_.resize(maxLevel + 1);
+
+    int cn = prevImg.channels();
+
+    if (cn == 1 || cn == 4)
+    {
+        prevImg.convertTo(prevPyr_[0], CV_32F);
+        nextImg.convertTo(nextPyr_[0], CV_32F);
+    }
+    else
+    {
+        buf_.resize(1);
+
+        cvtColor(prevImg, buf_[0], COLOR_BGR2BGRA);
+        buf_[0].convertTo(prevPyr_[0], CV_32F);
+
+        cvtColor(nextImg, buf_[0], COLOR_BGR2BGRA);
+        buf_[0].convertTo(nextPyr_[0], CV_32F);
+    }
+
+    for (int level = 1; level <= maxLevel; ++level)
+    {
+        pyrDown(prevPyr_[level - 1], prevPyr_[level]);
+        pyrDown(nextPyr_[level - 1], nextPyr_[level]);
+    }
+
+    //--------------------------------------------------------------------------
+    // Multithreading support
+    int index = -1;
+
+    do
+    {
+        unique_lock<tbb::mutex> ul(s_PyrLKOpticalFlow_Mutex);
+        for (unsigned int uiI = 0; uiI < 5; ++uiI)
+        {
+            if (index_vector_use[uiI])
+            {
+                index = uiI;
+                index_vector_use[uiI] = false;
+                break;
+            }
+        }
+        if (index < 0)
+            s_PyrLKOpticalFlow_ConditionVariable.wait(ul);
+
+        ul.unlock();
+    }while (index < 0);
+
+    //--------------------------------------------------------------------------
+
+    pyrlk::loadConstants_multi(make_int2(winSize.width, winSize.height), iters, index);
+
+    for (int level = maxLevel; level >= 0; level--)
+    {
+        if (cn == 1)
+        {
+            pyrlk::sparse1_multi(prevPyr_[level], nextPyr_[level],
+                prevPts.ptr<float2>(), nextPts.ptr<float2>(), status.ptr(),
+                level == 0 && err ? err->ptr<float>() : 0, prevPts.cols,
+                level, block, patch, StreamAccessor::getStream(stream), index);
+        }
+        else
+        {
+            pyrlk::sparse4_multi(prevPyr_[level], nextPyr_[level],
+                prevPts.ptr<float2>(), nextPts.ptr<float2>(), status.ptr(),
+                level == 0 && err ? err->ptr<float>() : 0, prevPts.cols,
+                level, block, patch, StreamAccessor::getStream(stream), index);
+        }
+    }
+
+    unique_lock<tbb::mutex> ul(s_PyrLKOpticalFlow_Mutex);
+    index_vector_use[index] = true;
+    s_PyrLKOpticalFlow_ConditionVariable.notify_one();
+}
+#else
+void cv::gpu::PyrLKOpticalFlow::sparse_multi(const GpuMat& /*prevImg*/,
+        const GpuMat& /*nextImg*/, const GpuMat& /*prevPts*/, GpuMat& /*nextPts*/,
+        GpuMat& /*status*/, Stream& /*stream*/, GpuMat* /*err*/)
+{
+    throw_notbb();
+}
+#endif
+
 void cv::gpu::PyrLKOpticalFlow::dense(const GpuMat& prevImg, const GpuMat& nextImg, GpuMat& u, GpuMat& v, GpuMat* err)
 {
     CV_Assert(prevImg.type() == CV_8UC1);

modules/gpu/test/test_optflow.cpp

@@ -44,6 +44,10 @@
 
 #ifdef HAVE_CUDA
 
+#ifdef HAVE_TBB
+#include <tbb/tbb.h>
+#endif
+
 using namespace cvtest;
 
 //////////////////////////////////////////////////////
@@ -322,6 +326,134 @@ GPU_TEST_P(PyrLKOpticalFlow, Sparse)
     ASSERT_LE(bad_ratio, 0.01);
 }
 
+#ifdef HAVE_TBB
+
+struct Sparse_Multi_Functor
+{
+    explicit Sparse_Multi_Functor(const cv::Mat& in_frame0, const cv::Mat& in_frame1,
+                                  const cv::Mat& in_pts_mat,
+                                  cv::gpu::GpuMat* in_d_pts,
+                                  cv::gpu::GpuMat* in_d_nextPts,
+                                  cv::gpu::GpuMat* in_d_status,
+                                  cv::gpu::Stream* in_streams):
+                                          m_frame0(in_frame0), m_frame1(in_frame1),
+                                          m_pts_mat(in_pts_mat),
+                                          m_d_pts(in_d_pts), m_d_nextPts(in_d_nextPts),
+                                          m_d_status(in_d_status), m_streams(in_streams){}
+
+    void operator()( const tbb::blocked_range<size_t>& r ) const
+    {
+        for( size_t i = r.begin(); i != r.end(); ++i )
+        {
+            m_d_pts[i].upload(m_pts_mat);
+            cv::gpu::PyrLKOpticalFlow pyrLK;
+            pyrLK.sparse_multi(loadMat(m_frame0), loadMat(m_frame1), m_d_pts[i],
+                               m_d_nextPts[i], m_d_status[i], m_streams[i]);
+            m_streams[i].waitForCompletion();
+        }
+    }
+
+    const cv::Mat& m_frame0;
+    const cv::Mat& m_frame1;
+    const cv::Mat& m_pts_mat;
+    cv::gpu::GpuMat* m_d_pts;
+    cv::gpu::GpuMat* m_d_nextPts;
+    cv::gpu::GpuMat* m_d_status;
+    cv::gpu::Stream* m_streams;
+};
+
+
+GPU_TEST_P(PyrLKOpticalFlow, Sparse_Multi)
+{
+    cv::Mat frame0 = readImage("opticalflow/frame0.png", useGray ? cv::IMREAD_GRAYSCALE : cv::IMREAD_COLOR);
+    ASSERT_FALSE(frame0.empty());
+
+    cv::Mat frame1 = readImage("opticalflow/frame1.png", useGray ? cv::IMREAD_GRAYSCALE : cv::IMREAD_COLOR);
+    ASSERT_FALSE(frame1.empty());
+
+    cv::Mat gray_frame;
+    if (useGray)
+        gray_frame = frame0;
+    else
+        cv::cvtColor(frame0, gray_frame, cv::COLOR_BGR2GRAY);
+
+    std::vector<cv::Point2f> pts;
+    cv::goodFeaturesToTrack(gray_frame, pts, 1000, 0.01, 0.0);
+
+    //--------------------------------------------------------------------------
+    // GPU
+    const unsigned int NB_EXEC_LINES = 27;
+
+    cv::gpu::GpuMat d_pts[NB_EXEC_LINES];
+    cv::gpu::GpuMat d_nextPts[NB_EXEC_LINES];
+    cv::gpu::GpuMat d_status[NB_EXEC_LINES];
+    cv::gpu::Stream streams[NB_EXEC_LINES];
+
+    cv::Mat pts_mat(1, (int) pts.size(), CV_32FC2, (void*) &pts[0]);
+
+    tbb::parallel_for(tbb::blocked_range<size_t>(0, NB_EXEC_LINES),
+                      Sparse_Multi_Functor(frame0, frame1, pts_mat,
+                                           d_pts, d_nextPts, d_status, streams));
+
+    std::vector<cv::Point2f> nextPts[NB_EXEC_LINES];
+    std::vector<unsigned char> status[NB_EXEC_LINES];
+
+    for (unsigned int i = 0; i < NB_EXEC_LINES; ++i)
+    {
+        nextPts[i].resize(d_nextPts[i].cols);
+        cv::Mat nextPts_mat(1, d_nextPts[i].cols, CV_32FC2, (void*) &(nextPts[i][0]));
+        d_nextPts[i].download(nextPts_mat);
+
+        status[i].resize(d_status[i].cols);
+        cv::Mat status_mat(1, d_status[i].cols, CV_8UC1, (void*) &(status[i][0]));
+        d_status[i].download(status_mat);
+    }
+
+    //--------------------------------------------------------------------------
+    // CPU
+    std::vector<cv::Point2f> nextPts_gold;
+    std::vector<unsigned char> status_gold;
+    cv::calcOpticalFlowPyrLK(frame0, frame1, pts, nextPts_gold, status_gold, cv::noArray());
+
+    //--------------------------------------------------------------------------
+    // CHECKS
+    for (unsigned int uiI = 0; uiI < NB_EXEC_LINES; ++uiI)
+    {
+        ASSERT_EQ(nextPts_gold.size(), nextPts[uiI].size());
+        ASSERT_EQ(status_gold.size(), status[uiI].size());
+    }
+
+    size_t mistmatch = 0;
+    for (unsigned int uiI = 0; uiI < NB_EXEC_LINES; ++uiI)
+    {
+        for (size_t i = 0; i < nextPts[uiI].size(); ++i)
+        {
+            cv::Point2i a = nextPts[uiI][i];
+            cv::Point2i b = nextPts_gold[i];
+
+            if (status[uiI][i] != status_gold[i])
+            {
+                ++mistmatch;
+                continue;
+            }
+
+            if (status[uiI][i])
+            {
+                bool eq = std::abs(a.x - b.x) <= 1 && std::abs(a.y - b.y) <= 1;
+
+                if (!eq)
+                    ++mistmatch;
+            }
+        }
+    }
+
+    double bad_ratio = static_cast<double>(mistmatch) / (nextPts[0].size() * NB_EXEC_LINES);
+
+    ASSERT_LE(bad_ratio, 0.01);
+}
+
+#endif // HAVE_TBB
+
 INSTANTIATE_TEST_CASE_P(GPU_Video, PyrLKOpticalFlow, testing::Combine(
     ALL_DEVICES,
     testing::Values(UseGray(true), UseGray(false))));

modules/highgui/src/cap_dc1394_v2.cpp

@@ -192,7 +192,11 @@ CvDC1394::~CvDC1394()
     dc = 0;
 }
 
-static CvDC1394 dc1394;
+static CvDC1394& getDC1394()
+{
+    static CvDC1394 dc1394;
+    return dc1394;
+}
 
 class CvCaptureCAM_DC1394_v2_CPP : public CvCapture
 {
@@ -451,7 +455,7 @@ bool CvCaptureCAM_DC1394_v2_CPP::startCapture()
     code = dc1394_capture_setup(dcCam, nDMABufs, DC1394_CAPTURE_FLAGS_DEFAULT);
     if (code >= 0)
     {
-        FD_SET(dc1394_capture_get_fileno(dcCam), &dc1394.camFds);
+        FD_SET(dc1394_capture_get_fileno(dcCam), &getDC1394().camFds);
         dc1394_video_set_transmission(dcCam, DC1394_ON);
         if (cameraId == VIDERE)
         {
@@ -477,15 +481,15 @@ bool CvCaptureCAM_DC1394_v2_CPP::open(int index)
 
     close();
 
-    if (!dc1394.dc)
+    if (!getDC1394().dc)
         goto _exit_;
 
-    err = dc1394_camera_enumerate(dc1394.dc, &cameraList);
+    err = dc1394_camera_enumerate(getDC1394().dc, &cameraList);
     if (err < 0 || !cameraList || (unsigned)index >= (unsigned)cameraList->num)
         goto _exit_;
 
     guid = cameraList->ids[index].guid;
-    dcCam = dc1394_camera_new(dc1394.dc, guid);
+    dcCam = dc1394_camera_new(getDC1394().dc, guid);
     if (!dcCam)
         goto _exit_;
 
@@ -510,8 +514,8 @@ void CvCaptureCAM_DC1394_v2_CPP::close()
         // check for fileno valid before using
         int fileno=dc1394_capture_get_fileno(dcCam);
 
-        if (fileno>=0 && FD_ISSET(fileno, &dc1394.camFds))
+        if (fileno>=0 && FD_ISSET(fileno, &getDC1394().camFds))
-            FD_CLR(fileno, &dc1394.camFds);
+            FD_CLR(fileno, &getDC1394().camFds);
         dc1394_video_set_transmission(dcCam, DC1394_OFF);
         dc1394_capture_stop(dcCam);
         dc1394_camera_free(dcCam);

modules/highgui/src/cap_ffmpeg_impl.hpp

@@ -173,6 +173,10 @@ extern "C" {
 #define AV_PIX_FMT_GRAY16BE PIX_FMT_GRAY16BE
 #endif
 
+#ifndef PKT_FLAG_KEY
+#define PKT_FLAG_KEY AV_PKT_FLAG_KEY
+#endif
+
 #if LIBAVUTIL_BUILD >= (LIBAVUTIL_VERSION_MICRO >= 100 \
     ? CALC_FFMPEG_VERSION(52, 38, 100) : CALC_FFMPEG_VERSION(52, 13, 0))
 #define USE_AV_FRAME_GET_BUFFER 1
@@ -1481,13 +1485,20 @@ static AVStream *icv_add_video_stream_FFMPEG(AVFormatContext *oc,
     // some formats want stream headers to be seperate
     if(oc->oformat->flags & AVFMT_GLOBALHEADER)
     {
+#if LIBAVCODEC_BUILD > CALC_FFMPEG_VERSION(56, 35, 0)
+        c->flags |= AV_CODEC_FLAG_GLOBAL_HEADER;
+#else
         c->flags |= CODEC_FLAG_GLOBAL_HEADER;
+#endif
     }
 #endif
 
 #if LIBAVCODEC_BUILD >= CALC_FFMPEG_VERSION(52, 42, 0)
     st->avg_frame_rate = (AVRational){frame_rate, frame_rate_base};
 #endif
+#if LIBAVFORMAT_BUILD >= CALC_FFMPEG_VERSION(55, 20, 0)
+    st->time_base = c->time_base;
+#endif
 
     return st;
 }
@@ -1509,23 +1520,24 @@ static int icv_av_write_frame_FFMPEG( AVFormatContext * oc, AVStream * video_st,
 #endif
     int ret = OPENCV_NO_FRAMES_WRITTEN_CODE;
 
-    if (oc->oformat->flags & AVFMT_RAWPICTURE) {
+#if LIBAVFORMAT_BUILD < CALC_FFMPEG_VERSION(57, 0, 0)
+    if (oc->oformat->flags & AVFMT_RAWPICTURE)
+    {
         /* raw video case. The API will change slightly in the near
            futur for that */
         AVPacket pkt;
         av_init_packet(&pkt);
 
-#ifndef PKT_FLAG_KEY
-#define PKT_FLAG_KEY AV_PKT_FLAG_KEY
-#endif
-
         pkt.flags |= PKT_FLAG_KEY;
         pkt.stream_index= video_st->index;
         pkt.data= (uint8_t *)picture;
         pkt.size= sizeof(AVPicture);
 
         ret = av_write_frame(oc, &pkt);
-    } else {
+    }
+    else
+#endif
+    {
         /* encode the image */
         AVPacket pkt;
         av_init_packet(&pkt);
@@ -1683,7 +1695,9 @@ void CvVideoWriter_FFMPEG::close()
     /* write the trailer, if any */
     if(ok && oc)
     {
-        if( (oc->oformat->flags & AVFMT_RAWPICTURE) == 0 )
+#if LIBAVFORMAT_BUILD < CALC_FFMPEG_VERSION(57, 0, 0)
+        if (!(oc->oformat->flags & AVFMT_RAWPICTURE))
+#endif
         {
             for(;;)
             {
@@ -1917,7 +1931,10 @@ bool CvVideoWriter_FFMPEG::open( const char * filename, int fourcc,
 
     outbuf = NULL;
 
-    if (!(oc->oformat->flags & AVFMT_RAWPICTURE)) {
+#if LIBAVFORMAT_BUILD < CALC_FFMPEG_VERSION(57, 0, 0)
+    if (!(oc->oformat->flags & AVFMT_RAWPICTURE))
+#endif
+    {
         /* allocate output buffer */
         /* assume we will never get codec output with more than 4 bytes per pixel... */
         outbuf_size = width*height*4;
@@ -2211,7 +2228,11 @@ AVStream* OutputMediaStream_FFMPEG::addVideoStream(AVFormatContext *oc, CV_CODEC
         // some formats want stream headers to be seperate
         if (oc->oformat->flags & AVFMT_GLOBALHEADER)
         {
-            c->flags |= CODEC_FLAG_GLOBAL_HEADER;
+            #if LIBAVCODEC_BUILD > CALC_FFMPEG_VERSION(56, 35, 0)
+                c->flags |= AV_CODEC_FLAG_GLOBAL_HEADER;
+            #else
+                c->flags |= CODEC_FLAG_GLOBAL_HEADER;
+            #endif
         }
     #endif
 

modules/highgui/src/cap_qt.cpp

@@ -634,11 +634,11 @@ static int icvOpenCamera_QT (CvCapture_QT_Cam * capture, const int index)
     OPENCV_ASSERT (capture,            "icvOpenCamera_QT", "'capture' is a NULL-pointer");
     OPENCV_ASSERT (index >=0, "icvOpenCamera_QT", "camera index is negative");
 
-    ComponentDescription	component_description;
+    ComponentDescription    component_description;
-    Component				component = 0;
+    Component               component = 0;
     int                     number_of_inputs = 0;
     Rect                    myRect;
-    ComponentResult			result = noErr;
+    ComponentResult         result = noErr;
 
 
     // travers all components and count video digitizer channels

modules/highgui/test/test_ffmpeg.cpp

@@ -320,6 +320,7 @@ public:
 
             for (unsigned int i = 0; i < frameCount && next; ++i)
             {
+                SCOPED_TRACE(cv::format("frame=%d", (int)i));
                 Mat actual;
                 (*capture) >> actual;
 

modules/stitching/src/matchers.cpp

@@ -543,6 +543,7 @@ void FeaturesMatcher::operator ()(const vector<ImageFeatures> &features, vector<
             if (features[i].keypoints.size() > 0 && features[j].keypoints.size() > 0 && mask_(i, j))
                 near_pairs.push_back(make_pair(i, j));
 
+    pairwise_matches.clear(); // clear history values
     pairwise_matches.resize(num_images * num_images);
     MatchPairsBody body(*this, features, pairwise_matches, near_pairs);
 

samples/gpu/CMakeLists.txt

@@ -90,6 +90,12 @@ if(BUILD_EXAMPLES AND OCV_DEPENDENCIES_FOUND)
     list(REMOVE_ITEM all_samples "driver_api_multi.cpp")
     list(REMOVE_ITEM all_samples "driver_api_stereo_multi.cpp")
   endif()
+  if(NOT HAVE_CUDA
+      OR NOT HAVE_TBB
+      OR OpenCV_FOUND  # via find_package() there is no access to cvconfig.h
+  )
+    list(REMOVE_ITEM all_samples "pyrlk_optical_flow_multithreading.cpp")
+  endif()
 
   foreach(sample_filename ${all_samples})
     get_filename_component(sample ${sample_filename} NAME_WE)
@@ -111,6 +117,9 @@ if (OCV_DEPENDENCIES_FOUND AND INSTALL_C_EXAMPLES AND NOT WIN32)
     list_filterout(install_list ".*driver_api_multi.cpp")
     list_filterout(install_list ".*driver_api_stereo_multi.cpp")
   endif()
+  if(NOT HAVE_CUDA OR NOT HAVE_TBB)
+      list(REMOVE_ITEM install_list "pyrlk_optical_flow_multithreading.cpp")
+  endif()
   install(FILES ${install_list}
           DESTINATION "${OPENCV_SAMPLES_SRC_INSTALL_PATH}/gpu"
           PERMISSIONS OWNER_READ OWNER_WRITE GROUP_READ WORLD_READ COMPONENT samples)

samples/gpu/pyrlk_optical_flow_multithreading.cpp

@@ -0,0 +1,272 @@
+#include <iostream>
+#include <vector>
+#include <sstream>
+
+#include "opencv2/core/core.hpp"
+#include "cvconfig.h"
+
+#ifdef HAVE_TBB
+#include <tbb/parallel_for_each.h>
+#include <tbb/task_scheduler_init.h>
+
+#include "opencv2/imgproc/imgproc.hpp"
+#include "opencv2/highgui/highgui.hpp"
+#include "opencv2/video/video.hpp"
+#include "opencv2/gpu/gpu.hpp"
+
+using namespace std;
+using namespace cv;
+using namespace cv::gpu;
+
+static void download(const GpuMat& d_mat, vector<Point2f>& vec)
+{
+    vec.resize(d_mat.cols);
+    Mat mat(1, d_mat.cols, CV_32FC2, (void*)&vec[0]);
+    d_mat.download(mat);
+}
+
+static void download(const GpuMat& d_mat, vector<uchar>& vec)
+{
+    vec.resize(d_mat.cols);
+    Mat mat(1, d_mat.cols, CV_8UC1, (void*)&vec[0]);
+    d_mat.download(mat);
+}
+
+static void drawArrows(Mat& frame, const vector<Point2f>& prevPts, const vector<Point2f>& nextPts, const vector<uchar>& status, Scalar line_color = Scalar(0, 0, 255))
+{
+    for (size_t i = 0; i < prevPts.size(); ++i)
+    {
+        if (status[i])
+        {
+            int line_thickness = 1;
+
+            Point p = prevPts[i];
+            Point q = nextPts[i];
+
+            double angle = atan2((double) p.y - q.y, (double) p.x - q.x);
+
+            double hypotenuse = sqrt( (double)(p.y - q.y)*(p.y - q.y) + (double)(p.x - q.x)*(p.x - q.x) );
+
+            if (hypotenuse < 1.0)
+                continue;
+
+            // Here we lengthen the arrow by a factor of three.
+            q.x = (int) (p.x - 3 * hypotenuse * cos(angle));
+            q.y = (int) (p.y - 3 * hypotenuse * sin(angle));
+
+            // Now we draw the main line of the arrow.
+            line(frame, p, q, line_color, line_thickness);
+
+            // Now draw the tips of the arrow. I do some scaling so that the
+            // tips look proportional to the main line of the arrow.
+
+            p.x = (int) (q.x + 9 * cos(angle + CV_PI / 4));
+            p.y = (int) (q.y + 9 * sin(angle + CV_PI / 4));
+            line(frame, p, q, line_color, line_thickness);
+
+            p.x = (int) (q.x + 9 * cos(angle - CV_PI / 4));
+            p.y = (int) (q.y + 9 * sin(angle - CV_PI / 4));
+            line(frame, p, q, line_color, line_thickness);
+        }
+    }
+}
+
+template <typename T> inline T clamp (T x, T a, T b)
+{
+    return ((x) > (a) ? ((x) < (b) ? (x) : (b)) : (a));
+}
+
+template <typename T> inline T mapValue(T x, T a, T b, T c, T d)
+{
+    x = clamp(x, a, b);
+    return c + (d - c) * (x - a) / (b - a);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+////////////////////////////////////////////////////////////////////////////////
+struct S_Thread_data
+{
+    Size winSize;
+    int maxLevel;
+    int iters;
+
+    Stream stream;
+    Mat frame0;
+    Mat frame1;
+    Mat frame1Gray;
+
+    GpuMat d_frame0Gray;
+    GpuMat d_prevPts;
+    bool useGray;
+};
+
+struct pyrLK_task
+{
+    pyrLK_task(size_t n):
+        _n(n),
+        _thread_data(NULL){}
+
+    void operator()()
+    {
+        // Sparse
+
+        PyrLKOpticalFlow d_pyrLK;
+
+        d_pyrLK.winSize.width = _thread_data->winSize.width;
+        d_pyrLK.winSize.height = _thread_data->winSize.height;
+        d_pyrLK.maxLevel = _thread_data->maxLevel;
+        d_pyrLK.iters = _thread_data->iters;
+
+        GpuMat d_frame0(_thread_data->frame0);
+        GpuMat d_frame1(_thread_data->frame1);
+        GpuMat d_frame1Gray(_thread_data->frame1Gray);
+        GpuMat d_nextPts;
+        GpuMat d_status;
+
+        bool useGray = _thread_data->useGray;
+
+        d_pyrLK.sparse_multi(useGray ? _thread_data->d_frame0Gray : d_frame0,
+                    useGray ? d_frame1Gray : d_frame1,
+                            _thread_data->d_prevPts, d_nextPts,
+                    d_status, _thread_data->stream, NULL);
+
+        // Draw arrows
+
+        vector<Point2f> prevPts(_thread_data->d_prevPts.cols);
+        download(_thread_data->d_prevPts, prevPts);
+
+        vector<Point2f> nextPts(d_nextPts.cols);
+        download(d_nextPts, nextPts);
+
+        vector<uchar> status(d_status.cols);
+        download(d_status, status);
+
+        drawArrows(_thread_data->frame0, prevPts, nextPts, status, Scalar(255, 0, 0));
+    }
+
+    size_t _n;
+    struct S_Thread_data* _thread_data;
+};
+
+template <typename T> struct invoker {
+  void operator()(T& it) const {it();}
+};
+
+#define THREADS_NB    12
+
+int main(int argc, const char* argv[])
+{
+    const char* keys =
+        "{ h            | help           | false | print help message }"
+        "{ l            | left           |       | specify left image }"
+        "{ r            | right          |       | specify right image }"
+        "{ gray         | gray           | false | use grayscale sources [PyrLK Sparse] }"
+        "{ win_size     | win_size       | 21    | specify windows size [PyrLK] }"
+        "{ max_level    | max_level      | 3     | specify max level [PyrLK] }"
+        "{ iters        | iters          | 30    | specify iterations count [PyrLK] }"
+        "{ points       | points         | 4000  | specify points count [GoodFeatureToTrack] }"
+        "{ min_dist     | min_dist       | 0     | specify minimal distance between points [GoodFeatureToTrack] }";
+
+    CommandLineParser cmd(argc, argv, keys);
+
+    if (cmd.get<bool>("help"))
+    {
+        cout << "Usage: pyrlk_optical_flow_multithreading [options]" << endl;
+        cout << "Avaible options:" << endl;
+        cmd.printParams();
+        return 0;
+    }
+
+    string fname0 = cmd.get<string>("left");
+    string fname1 = cmd.get<string>("right");
+
+    if (fname0.empty() || fname1.empty())
+    {
+        cerr << "Missing input file names" << endl;
+        return -1;
+    }
+
+    bool useGray = cmd.get<bool>("gray");
+    int winSize = cmd.get<int>("win_size");
+    int maxLevel = cmd.get<int>("max_level");
+    int iters = cmd.get<int>("iters");
+    int points = cmd.get<int>("points");
+    double minDist = cmd.get<double>("min_dist");
+
+    Mat frame0 = imread(fname0);
+    Mat frame1 = imread(fname1);
+
+    if (frame0.empty() || frame1.empty())
+    {
+        cout << "Can't load input images" << endl;
+        return -1;
+    }
+
+    cout << "Image size : " << frame0.cols << " x " << frame0.rows << endl;
+    cout << "Points count : " << points << endl;
+
+    cout << endl;
+
+    Mat frame0Gray;
+    cvtColor(frame0, frame0Gray, COLOR_BGR2GRAY);
+    Mat frame1Gray;
+    cvtColor(frame1, frame1Gray, COLOR_BGR2GRAY);
+
+    // goodFeaturesToTrack
+
+    GoodFeaturesToTrackDetector_GPU detector(points, 0.01, minDist);
+
+    GpuMat d_frame0Gray(frame0Gray);
+    GpuMat d_prevPts;
+
+    detector(d_frame0Gray, d_prevPts);
+
+    // Sparse
+
+    tbb::task_scheduler_init init(THREADS_NB);
+
+    std::vector<pyrLK_task> tasks;
+
+    S_Thread_data s_thread_data[THREADS_NB];
+
+    for (unsigned int uiI = 0; uiI < THREADS_NB; ++uiI)
+    {
+        s_thread_data[uiI].stream = Stream();
+        s_thread_data[uiI].frame0 = frame0.clone();
+        s_thread_data[uiI].frame1 = frame1.clone();
+        s_thread_data[uiI].frame1Gray = frame0Gray.clone();
+
+        s_thread_data[uiI].iters = iters;
+        s_thread_data[uiI].useGray = useGray;
+        s_thread_data[uiI].maxLevel = maxLevel;
+        s_thread_data[uiI].winSize.height = winSize;
+        s_thread_data[uiI].winSize.width = winSize;
+        s_thread_data[uiI].d_frame0Gray = d_frame0Gray.clone();
+        s_thread_data[uiI].d_prevPts = d_prevPts.clone();
+
+        tasks.push_back(pyrLK_task(uiI));
+        tasks.back()._thread_data = &(s_thread_data[uiI]);
+    }
+
+    tbb::parallel_for_each(tasks.begin(),tasks.end(),invoker<pyrLK_task>());
+
+    for (unsigned int uiI = 0; uiI < THREADS_NB; ++uiI)
+    {
+        stringstream ss;
+        ss << "PyrLK MultiThreading [Sparse] " << uiI;
+
+        imshow(ss.str(), s_thread_data[uiI].frame0);
+        ss.str("");
+    }
+
+    waitKey();
+
+    return 0;
+}
+#else
+int main(int , const char* [])
+{
+    std::cout << "This example pyrlk_optical_flow_multithreading must be compiled with TBB Option" << std::endl;
+    return 0;
+}
+#endif // HAVE_TBB