From: Vladislav Vinogradov <no@email>
Date: Mon, 31 Jan 2011 13:20:52 +0000 (+0000)
Subject: fixed gpu tests (BruteForceMatcher_GPU, divide, phase, cartToPolar, async)
X-Git-Tag: accepted/2.0/20130307.220821~3585
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=8274ed22e42c7b7c0ff6c05708208a3b6cb75259;p=profile%2Fivi%2Fopencv.git

fixed gpu tests (BruteForceMatcher_GPU, divide, phase, cartToPolar, async)
minor code refactoring
---

diff --git a/modules/gpu/include/opencv2/gpu/gpu.hpp b/modules/gpu/include/opencv2/gpu/gpu.hpp
index b719417..2531d1e 100644
--- a/modules/gpu/include/opencv2/gpu/gpu.hpp
+++ b/modules/gpu/include/opencv2/gpu/gpu.hpp
@@ -671,10 +671,12 @@ namespace cv
         //! output will have CV_32FC1 type
         CV_EXPORTS void rectStdDev(const GpuMat& src, const GpuMat& sqr, GpuMat& dst, const Rect& rect);
 
-        //! applies Canny edge detector and produces the edge map
-        //! supprots only CV_8UC1 source type
-        //! disabled until fix crash
-        CV_EXPORTS void Canny(const GpuMat& image, GpuMat& edges, double threshold1, double threshold2, int apertureSize = 3);
+        // applies Canny edge detector and produces the edge map
+        // disabled until fix crash
+        //CV_EXPORTS void Canny(const GpuMat& image, GpuMat& edges, double threshold1, double threshold2, int apertureSize = 3);
+        //CV_EXPORTS void Canny(const GpuMat& image, GpuMat& edges, GpuMat& buffer, double threshold1, double threshold2, int apertureSize = 3);
+        //CV_EXPORTS void Canny(const GpuMat& srcDx, const GpuMat& srcDy, GpuMat& edges, double threshold1, double threshold2, int apertureSize = 3);
+        //CV_EXPORTS void Canny(const GpuMat& srcDx, const GpuMat& srcDy, GpuMat& edges, GpuMat& buffer, double threshold1, double threshold2, int apertureSize = 3);
 
         //! computes Harris cornerness criteria at each image pixel
         CV_EXPORTS void cornerHarris(const GpuMat& src, GpuMat& dst, int blockSize, int ksize, double k, int borderType=BORDER_REFLECT101);
diff --git a/modules/gpu/src/brute_force_matcher.cpp b/modules/gpu/src/brute_force_matcher.cpp
index 3447121..4806e67 100644
--- a/modules/gpu/src/brute_force_matcher.cpp
+++ b/modules/gpu/src/brute_force_matcher.cpp
@@ -104,6 +104,18 @@ namespace cv { namespace gpu { namespace bfmatcher
         const DevMem2D& mask, const DevMem2Di& trainIdx, unsigned int* nMatches, const DevMem2Df& distance);
 }}}
 
+namespace
+{
+    class ImgIdxSetter
+    {
+    public:
+        ImgIdxSetter(int imgIdx_) : imgIdx(imgIdx_) {}
+        void operator()(DMatch& m) const {m.imgIdx = imgIdx;}
+    private:
+        int imgIdx;
+    };
+}
+
 cv::gpu::BruteForceMatcher_GPU_base::BruteForceMatcher_GPU_base(DistType distType_) : distType(distType_)
 {
 }
@@ -185,7 +197,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::matchDownload(const GpuMat& trainIdx,
         return;
 
     CV_Assert(trainIdx.type() == CV_32SC1 && trainIdx.isContinuous());
-    CV_Assert(distance.type() == CV_32FC1 && distance.isContinuous() && distance.size().area() == trainIdx.size().area());
+    CV_Assert(distance.type() == CV_32FC1 && distance.isContinuous() && distance.cols == trainIdx.cols);
 
     const int nQuery = trainIdx.cols;
 
@@ -309,8 +321,8 @@ void cv::gpu::BruteForceMatcher_GPU_base::matchDownload(const GpuMat& trainIdx,
         return;
 
     CV_Assert(trainIdx.type() == CV_32SC1 && trainIdx.isContinuous());
-    CV_Assert(imgIdx.type() == CV_32SC1 && imgIdx.isContinuous());
-    CV_Assert(distance.type() == CV_32FC1 && distance.isContinuous());
+    CV_Assert(imgIdx.type() == CV_32SC1 && imgIdx.isContinuous() && imgIdx.cols == trainIdx.cols);
+    CV_Assert(distance.type() == CV_32FC1 && distance.isContinuous() && imgIdx.cols == trainIdx.cols);
 
     const int nQuery = trainIdx.cols;
 
@@ -390,7 +402,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::knnMatch(const GpuMat& queryDescs, con
     trainIdx.setTo(Scalar::all(-1));
     distance.create(nQuery, k, CV_32F);
 
-    allDist.create(nQuery, nTrain, CV_32F);
+    ensureSizeIsEnough(nQuery, nTrain, CV_32FC1, allDist);
 
     match_caller_t func = match_callers[distType][queryDescs.depth()];
     CV_Assert(func != 0);
@@ -451,18 +463,6 @@ void cv::gpu::BruteForceMatcher_GPU_base::knnMatch(const GpuMat& queryDescs, con
     knnMatchDownload(trainIdx, distance, matches, compactResult);
 }
 
-namespace
-{
-    class ImgIdxSetter
-    {
-    public:
-        ImgIdxSetter(int imgIdx_) : imgIdx(imgIdx_) {}
-        void operator()(DMatch& m) const {m.imgIdx = imgIdx;}
-    private:
-        int imgIdx;
-    };
-}
-
 void cv::gpu::BruteForceMatcher_GPU_base::knnMatch(const GpuMat& queryDescs,
     vector< vector<DMatch> >& matches, int knn, const vector<GpuMat>& masks, bool compactResult)
 {
@@ -538,9 +538,9 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatch(const GpuMat& queryDescs,
 
     CV_Assert(queryDescs.channels() == 1 && queryDescs.depth() < CV_64F);
     CV_Assert(trainDescs.type() == queryDescs.type() && trainDescs.cols == queryDescs.cols);
-    CV_Assert(trainIdx.empty() || trainIdx.rows == nQuery);
+    CV_Assert(trainIdx.empty() || (trainIdx.rows == nQuery && trainIdx.size() == distance.size()));
 
-    nMatches.create(1, nQuery, CV_32SC1);
+    ensureSizeIsEnough(1, nQuery, CV_32SC1, nMatches);
     nMatches.setTo(Scalar::all(0));
     if (trainIdx.empty())
     {
@@ -561,7 +561,7 @@ void cv::gpu::BruteForceMatcher_GPU_base::radiusMatchDownload(const GpuMat& trai
         return;
 
     CV_Assert(trainIdx.type() == CV_32SC1);
-    CV_Assert(nMatches.type() == CV_32SC1 && nMatches.isContinuous() && nMatches.size().area() == trainIdx.rows);
+    CV_Assert(nMatches.type() == CV_32SC1 && nMatches.isContinuous() && nMatches.cols >= trainIdx.rows);
     CV_Assert(distance.type() == CV_32FC1 && distance.size() == trainIdx.size());
 
     const int nQuery = trainIdx.rows;
diff --git a/modules/gpu/src/cuda/brute_force_matcher.cu b/modules/gpu/src/cuda/brute_force_matcher.cu
index b28aee1..44f823d 100644
--- a/modules/gpu/src/cuda/brute_force_matcher.cu
+++ b/modules/gpu/src/cuda/brute_force_matcher.cu
@@ -64,6 +64,7 @@ namespace cv { namespace gpu { namespace bfmatcher
         {            
             return mask.ptr(queryIdx)[trainIdx] != 0;
         }
+
     private:
         PtrStep mask;
     };
@@ -82,6 +83,7 @@ namespace cv { namespace gpu { namespace bfmatcher
         {            
             return curMask.data == 0 || curMask.ptr(queryIdx)[trainIdx] != 0;
         }
+
     private:
         PtrStep* maskCollection;
         PtrStep curMask;
@@ -102,172 +104,55 @@ namespace cv { namespace gpu { namespace bfmatcher
     ///////////////////////////////////////////////////////////////////////////////
     // Reduce Sum
     
-    template <int BLOCK_DIM_X>
-    __device__ void reduceSum(float* sdiff, float mySum, int tid)
-    {
-        sdiff[tid] = mySum;
-        __syncthreads();
+    template <int BLOCK_DIM_X> __device__ void reduceSum(float* sdiff_row, float& mySum);
 
-        if (BLOCK_DIM_X == 512) 
-        {
-            if (tid < 256) 
-            { 
-                sdiff[tid] = mySum += sdiff[tid + 256]; __syncthreads(); 
-                sdiff[tid] = mySum += sdiff[tid + 128]; __syncthreads();
-                sdiff[tid] = mySum += sdiff[tid +  64]; __syncthreads();
-            }
-            volatile float* smem = sdiff;
-            smem[tid] = mySum += smem[tid + 32]; 
-            smem[tid] = mySum += smem[tid + 16]; 
-            smem[tid] = mySum += smem[tid +  8]; 
-            smem[tid] = mySum += smem[tid +  4]; 
-            smem[tid] = mySum += smem[tid +  2];
-            smem[tid] = mySum += smem[tid +  1]; 
-        }
-        if (BLOCK_DIM_X == 256)
-        {
-            if (tid < 128) 
-            { 
-                sdiff[tid] = mySum += sdiff[tid + 128]; __syncthreads(); 
-                sdiff[tid] = mySum += sdiff[tid +  64]; __syncthreads();
-            }
-            volatile float* smem = sdiff;
-            smem[tid] = mySum += smem[tid + 32]; 
-            smem[tid] = mySum += smem[tid + 16]; 
-            smem[tid] = mySum += smem[tid +  8]; 
-            smem[tid] = mySum += smem[tid +  4]; 
-            smem[tid] = mySum += smem[tid +  2];
-            smem[tid] = mySum += smem[tid +  1];
-        }
-        if (BLOCK_DIM_X == 128)
-        {
-            if (tid <  64) 
-            { 
-                sdiff[tid] = mySum += sdiff[tid +  64]; __syncthreads(); 
-            }
-            volatile float* smem = sdiff;
-            smem[tid] = mySum += smem[tid + 32]; 
-            smem[tid] = mySum += smem[tid + 16]; 
-            smem[tid] = mySum += smem[tid +  8]; 
-            smem[tid] = mySum += smem[tid +  4]; 
-            smem[tid] = mySum += smem[tid +  2];
-            smem[tid] = mySum += smem[tid +  1];
-        }
-        
-        volatile float* smem = sdiff;
-        if (BLOCK_DIM_X == 64) 
-        {
-            if (tid < 32) 
-            {
-                smem[tid] = mySum += smem[tid + 32]; 
-                smem[tid] = mySum += smem[tid + 16]; 
-                smem[tid] = mySum += smem[tid +  8]; 
-                smem[tid] = mySum += smem[tid +  4]; 
-                smem[tid] = mySum += smem[tid +  2];
-                smem[tid] = mySum += smem[tid +  1];  
-            }
-        }
-        if (BLOCK_DIM_X == 32) 
-        {
-            if (tid < 16) 
-            {
-                smem[tid] = mySum += smem[tid + 16]; 
-                smem[tid] = mySum += smem[tid +  8]; 
-                smem[tid] = mySum += smem[tid +  4]; 
-                smem[tid] = mySum += smem[tid +  2];
-                smem[tid] = mySum += smem[tid +  1];  
-            }
-        }
-        if (BLOCK_DIM_X == 16) 
-        {
-            if (tid < 8) 
-            {
-                smem[tid] = mySum += smem[tid +  8]; 
-                smem[tid] = mySum += smem[tid +  4]; 
-                smem[tid] = mySum += smem[tid +  2];
-                smem[tid] = mySum += smem[tid +  1];  
-            }
-        }
-        if (BLOCK_DIM_X == 8) 
-        {
-            if (tid < 4) 
-            {
-                smem[tid] = mySum += smem[tid +  4]; 
-                smem[tid] = mySum += smem[tid +  2];
-                smem[tid] = mySum += smem[tid +  1];  
-            }
-        }
-        if (BLOCK_DIM_X == 4) 
-        {
-            if (tid < 2) 
-            {
-                smem[tid] = mySum += smem[tid +  2];
-                smem[tid] = mySum += smem[tid +  1];  
-            }
-        }
-        if (BLOCK_DIM_X == 2) 
-        {
-            if (tid < 1) 
-            {
-                smem[tid] = mySum += smem[tid +  1];  
-            }
-        }
-    }
-
-    ///////////////////////////////////////////////////////////////////////////////
-    // loadDescsVals
-
-    template <int BLOCK_DIM_X, int BLOCK_DIM_Y, int MAX_DESCRIPTORS_LEN, typename T> 
-    __device__ void loadDescsVals(const T* descs, int desc_len, float* smem, float* queryVals)
+    template <> __device__ void reduceSum<16>(float* sdiff_row, float& mySum)
     {
-        const int tid = threadIdx.y * blockDim.x + threadIdx.x;
+        volatile float* smem = sdiff_row;
 
-        if (tid < desc_len)
+        smem[threadIdx.x] = mySum;
+        
+        if (threadIdx.x < 8) 
         {
-            smem[tid] = (float)descs[tid];
-        }
-        __syncthreads();
-
-        #pragma unroll
-        for (int i = threadIdx.x; i < MAX_DESCRIPTORS_LEN; i += BLOCK_DIM_X)
-        {
-            *queryVals = smem[i];
-            ++queryVals;
+            smem[threadIdx.x] = mySum += smem[threadIdx.x + 8]; 
+            smem[threadIdx.x] = mySum += smem[threadIdx.x + 4]; 
+            smem[threadIdx.x] = mySum += smem[threadIdx.x + 2];
+            smem[threadIdx.x] = mySum += smem[threadIdx.x + 1];  
         }
     }
 
     ///////////////////////////////////////////////////////////////////////////////
     // Distance
 
-    template <int BLOCK_DIM_X>
     class L1Dist
     {
     public:
-        __device__ L1Dist() : mySum(0) {}
+        __device__ L1Dist() : mySum(0.0f) {}
 
         __device__ void reduceIter(float val1, float val2)
         {
             mySum += fabs(val1 - val2);
         }
 
-        __device__ void reduceAll(float* sdiff, int tid)
+        template <int BLOCK_DIM_X>
+        __device__ void reduceAll(float* sdiff_row)
         {
-            reduceSum<BLOCK_DIM_X>(sdiff, mySum, tid);
+            reduceSum<BLOCK_DIM_X>(sdiff_row, mySum);
         }
 
-        static __device__ float finalResult(float res)
+        __device__ operator float() const
         {
-            return res;
+            return mySum;
         }
+
     private:
         float mySum;
     };
 
-    template <int BLOCK_DIM_X>
     class L2Dist
     {
     public:
-        __device__ L2Dist() : mySum(0) {}
+        __device__ L2Dist() : mySum(0.0f) {}
 
         __device__ void reduceIter(float val1, float val2)
         {
@@ -275,15 +160,17 @@ namespace cv { namespace gpu { namespace bfmatcher
             mySum += reg * reg;
         }
 
-        __device__ void reduceAll(float* sdiff, int tid)
+        template <int BLOCK_DIM_X>
+        __device__ void reduceAll(float* sdiff_row)
         {
-            reduceSum<BLOCK_DIM_X>(sdiff, mySum, tid);
+            reduceSum<BLOCK_DIM_X>(sdiff_row, mySum);
         }
 
-        static __device__ float finalResult(float res)
+        __device__ operator float() const
         {
-            return sqrtf(res);
+            return sqrtf(mySum);
         }
+
     private:
         float mySum;
     };
@@ -292,56 +179,81 @@ namespace cv { namespace gpu { namespace bfmatcher
     // reduceDescDiff
 
     template <int BLOCK_DIM_X, typename Dist, typename T> 
-    __device__ void reduceDescDiff(const T* queryDescs, const T* trainDescs, int desc_len, float* sdiff)
+    __device__ void reduceDescDiff(const T* queryDescs, const T* trainDescs, int desc_len, Dist& dist, 
+        float* sdiff_row)
     {
-        const int tid = threadIdx.x;
+        for (int i = threadIdx.x; i < desc_len; i += BLOCK_DIM_X)
+            dist.reduceIter(queryDescs[i], trainDescs[i]);
 
-        Dist dist;
+        dist.reduceAll<BLOCK_DIM_X>(sdiff_row);
+    }
 
-        for (int i = tid; i < desc_len; i += BLOCK_DIM_X)
-            dist.reduceIter(queryDescs[i], trainDescs[i]);
+///////////////////////////////////////////////////////////////////////////////////
+////////////////////////////////////// Match //////////////////////////////////////
+///////////////////////////////////////////////////////////////////////////////////
+    
+    ///////////////////////////////////////////////////////////////////////////////
+    // loadDescsVals
+
+    template <int BLOCK_DIM_X, int MAX_DESCRIPTORS_LEN, typename T> 
+    __device__ void loadDescsVals(const T* descs, int desc_len, float* queryVals, float* smem)
+    {
+        const int tid = threadIdx.y * blockDim.x + threadIdx.x;
+
+        if (tid < desc_len)
+        {
+            smem[tid] = (float)descs[tid];
+        }
+        __syncthreads();
 
-        dist.reduceAll(sdiff, tid);
+        #pragma unroll
+        for (int i = threadIdx.x; i < MAX_DESCRIPTORS_LEN; i += BLOCK_DIM_X)
+        {
+            *queryVals = smem[i];
+            ++queryVals;
+        }
     }
 
     ///////////////////////////////////////////////////////////////////////////////
-    // reduceDescDiff_smem
+    // reduceDescDiffCached
 
     template <int N> struct UnrollDescDiff
     {
         template <typename Dist, typename T>
-        static __device__ void calcCheck(Dist& dist, const float* queryVals, const T* trainDescs, 
-            int ind, int desc_len)
+        static __device__ void calcCheck(const float* queryVals, const T* trainDescs, int desc_len, 
+            Dist& dist, int ind)
         {
             if (ind < desc_len)
+            {
                 dist.reduceIter(*queryVals, trainDescs[ind]);
 
-            ++queryVals;
+                ++queryVals;
 
-            UnrollDescDiff<N - 1>::calcCheck(dist, queryVals, trainDescs, ind + blockDim.x, desc_len);
+                UnrollDescDiff<N - 1>::calcCheck(queryVals, trainDescs, desc_len, dist, ind + blockDim.x);
+            }
         }
 
         template <typename Dist, typename T>
-        static __device__ void calcWithoutCheck(Dist& dist, const float* queryVals, const T* trainDescs)
+        static __device__ void calcWithoutCheck(const float* queryVals, const T* trainDescs, Dist& dist)
         {
             dist.reduceIter(*queryVals, *trainDescs);
 
             ++queryVals;
             trainDescs += blockDim.x;
 
-            UnrollDescDiff<N - 1>::calcWithoutCheck(dist, queryVals, trainDescs);
+            UnrollDescDiff<N - 1>::calcWithoutCheck(queryVals, trainDescs, dist);
         }
     };
     template <> struct UnrollDescDiff<0>
     {
         template <typename Dist, typename T>
-        static __device__ void calcCheck(Dist& dist, const float* queryVals, const T* trainDescs, 
-            int ind, int desc_len)
+        static __device__ void calcCheck(const float* queryVals, const T* trainDescs, int desc_len, 
+            Dist& dist, int ind)
         {
         }
 
         template <typename Dist, typename T>
-        static __device__ void calcWithoutCheck(Dist& dist, const float* queryVals, const T* trainDescs)
+        static __device__ void calcWithoutCheck(const float* queryVals, const T* trainDescs, Dist& dist)
         {
         }
     };
@@ -351,106 +263,82 @@ namespace cv { namespace gpu { namespace bfmatcher
     struct DescDiffCalculator<BLOCK_DIM_X, MAX_DESCRIPTORS_LEN, false>
     {
         template <typename Dist, typename T>
-        static __device__ void calc(Dist& dist, const float* queryVals, const T* trainDescs, int desc_len)
+        static __device__ void calc(const float* queryVals, const T* trainDescs, int desc_len, Dist& dist)
         {
-            UnrollDescDiff<MAX_DESCRIPTORS_LEN / BLOCK_DIM_X>::calcCheck(dist, queryVals, trainDescs, 
-                threadIdx.x, desc_len);
+            UnrollDescDiff<MAX_DESCRIPTORS_LEN / BLOCK_DIM_X>::calcCheck(queryVals, trainDescs, desc_len, 
+                dist, threadIdx.x);
         }
     };
     template <int BLOCK_DIM_X, int MAX_DESCRIPTORS_LEN> 
     struct DescDiffCalculator<BLOCK_DIM_X, MAX_DESCRIPTORS_LEN, true>
     {
         template <typename Dist, typename T>
-        static __device__ void calc(Dist& dist, const float* queryVals, const T* trainDescs, int desc_len)
+        static __device__ void calc(const float* queryVals, const T* trainDescs, int desc_len, Dist& dist)
         {
-            UnrollDescDiff<MAX_DESCRIPTORS_LEN / BLOCK_DIM_X>::calcWithoutCheck(dist, queryVals, 
-                trainDescs + threadIdx.x);
+            UnrollDescDiff<MAX_DESCRIPTORS_LEN / BLOCK_DIM_X>::calcWithoutCheck(queryVals, 
+                trainDescs + threadIdx.x, dist);
         }
     };
 
     template <int BLOCK_DIM_X, int MAX_DESCRIPTORS_LEN, bool DESC_LEN_EQ_MAX_LEN, typename Dist, typename T>
-    __device__ void reduceDescDiff_smem(const float* queryVals, const T* trainDescs, int desc_len, float* sdiff)
-    {
-        const int tid = threadIdx.x;
+    __device__ void reduceDescDiffCached(const float* queryVals, const T* trainDescs, int desc_len, Dist& dist, 
+        float* sdiff_row)
+    {        
+        DescDiffCalculator<BLOCK_DIM_X, MAX_DESCRIPTORS_LEN, DESC_LEN_EQ_MAX_LEN>::calc(queryVals, 
+            trainDescs, desc_len, dist);
         
-        Dist dist;
-
-        DescDiffCalculator<BLOCK_DIM_X, MAX_DESCRIPTORS_LEN, DESC_LEN_EQ_MAX_LEN>::calc(dist, queryVals, 
-            trainDescs, desc_len);
-        
-        dist.reduceAll(sdiff, tid);
+        dist.reduceAll<BLOCK_DIM_X>(sdiff_row);
     }
 
-///////////////////////////////////////////////////////////////////////////////////
-////////////////////////////////////// Match //////////////////////////////////////
-///////////////////////////////////////////////////////////////////////////////////
-
     ///////////////////////////////////////////////////////////////////////////////
-    // warpReduceMin
+    // warpReduceMinIdxIdx
 
     template <int BLOCK_DIM_Y> 
-    __device__ void warpReduceMin(int tid, volatile float* sdata, volatile int* strainIdx, volatile int* simgIdx)
+    __device__ void warpReduceMinIdxIdx(float& myMin, int& myBestTrainIdx, int& myBestImgIdx, 
+        volatile float* sdata, volatile int* strainIdx, volatile int* simgIdx);
+
+    template <> 
+    __device__ void warpReduceMinIdxIdx<16>(float& myMin, int& myBestTrainIdx, int& myBestImgIdx, 
+        volatile float* smin, volatile int* strainIdx, volatile int* simgIdx)
     {
-        float minSum = sdata[tid];
+        const int tid = threadIdx.y * blockDim.x + threadIdx.x;
 
-        if (BLOCK_DIM_Y >= 64) 
-        {
-            float reg = sdata[tid + 32];
-            if (reg < minSum)
-            {
-                sdata[tid] = minSum = reg;
-                strainIdx[tid] = strainIdx[tid + 32];
-                simgIdx[tid] = simgIdx[tid + 32];
-            }
-        }
-        if (BLOCK_DIM_Y >= 32) 
+        if (tid < 8)
         {
-            float reg = sdata[tid + 16];
-            if (reg < minSum)
-            {
-                sdata[tid] = minSum = reg;
-                strainIdx[tid] = strainIdx[tid + 16];
-                simgIdx[tid] = simgIdx[tid + 16];
-            }
-        }
-        if (BLOCK_DIM_Y >= 16) 
-        {
-            float reg = sdata[tid + 8];
-            if (reg < minSum)
+            myMin = smin[tid];
+            myBestTrainIdx = strainIdx[tid];
+            myBestImgIdx = simgIdx[tid];
+
+            float reg = smin[tid + 8];
+            if (reg < myMin)
             {
-                sdata[tid] = minSum = reg;
-                strainIdx[tid] = strainIdx[tid + 8];
-                simgIdx[tid] = simgIdx[tid + 8];
+                smin[tid] = myMin = reg;
+                strainIdx[tid] = myBestTrainIdx = strainIdx[tid + 8];
+                simgIdx[tid] = myBestImgIdx = simgIdx[tid + 8];
             }
-        }
-        if (BLOCK_DIM_Y >= 8) 
-        { 
-            float reg = sdata[tid + 4];
-            if (reg < minSum)
+
+            reg = smin[tid + 4];
+            if (reg < myMin)
             {
-                sdata[tid] = minSum = reg;
-                strainIdx[tid] = strainIdx[tid + 4];
-                simgIdx[tid] = simgIdx[tid + 4];
+                smin[tid] = myMin = reg;
+                strainIdx[tid] = myBestTrainIdx = strainIdx[tid + 4];
+                simgIdx[tid] = myBestImgIdx = simgIdx[tid + 4];
             }
-        }
-        if (BLOCK_DIM_Y >= 4) 
-        { 
-            float reg = sdata[tid + 2];
-            if (reg < minSum)
+        
+            reg = smin[tid + 2];
+            if (reg < myMin)
             {
-                sdata[tid] = minSum = reg;
-                strainIdx[tid] = strainIdx[tid + 2];
-                simgIdx[tid] = simgIdx[tid + 2];
+                smin[tid] = myMin = reg;
+                strainIdx[tid] = myBestTrainIdx = strainIdx[tid + 2];
+                simgIdx[tid] = myBestImgIdx = simgIdx[tid + 2];
             }
-        }
-        if (BLOCK_DIM_Y >= 2) 
-        {
-            float reg = sdata[tid + 1];
-            if (reg < minSum)
+        
+            reg = smin[tid + 1];
+            if (reg < myMin)
             {
-                sdata[tid] = minSum = reg;
-                strainIdx[tid] = strainIdx[tid + 1];
-                simgIdx[tid] = simgIdx[tid + 1];
+                smin[tid] = myMin = reg;
+                strainIdx[tid] = myBestTrainIdx = strainIdx[tid + 1];
+                simgIdx[tid] = myBestImgIdx = simgIdx[tid + 1];
             }
         }
     }
@@ -458,9 +346,9 @@ namespace cv { namespace gpu { namespace bfmatcher
     ///////////////////////////////////////////////////////////////////////////////
     // findBestMatch
 
-    template <int BLOCK_DIM_Y, typename Dist>
-    __device__ void findBestMatch(int queryIdx, float myMin, int myBestTrainIdx, int myBestImgIdx, 
-        float* smin, int* strainIdx, int* simgIdx, int* trainIdx, int* imgIdx, float* distance)
+    template <int BLOCK_DIM_Y>
+    __device__ void findBestMatch(float& myMin, int& myBestTrainIdx, int& myBestImgIdx, 
+        float* smin, int* strainIdx, int* simgIdx)
     {
         if (threadIdx.x == 0)
         {
@@ -470,27 +358,13 @@ namespace cv { namespace gpu { namespace bfmatcher
         }
         __syncthreads();
 
-        const int tid = threadIdx.y * blockDim.x + threadIdx.x;
-
-        if (tid < 32)
-            warpReduceMin<BLOCK_DIM_Y>(tid, smin, strainIdx, simgIdx);
-
-        if (threadIdx.x == 0 && threadIdx.y == 0)
-        {
-            float minSum = smin[0];
-            int bestTrainIdx = strainIdx[0];
-            int bestImgIdx = simgIdx[0];
-
-            imgIdx[queryIdx] = bestImgIdx;
-            trainIdx[queryIdx] = bestTrainIdx;
-            distance[queryIdx] = Dist::finalResult(minSum);
-        }
+        warpReduceMinIdxIdx<BLOCK_DIM_Y>(myMin, myBestTrainIdx, myBestImgIdx, smin, strainIdx, simgIdx);
     }
     
     ///////////////////////////////////////////////////////////////////////////////
     // ReduceDescCalculator
 
-    template <int BLOCK_DIM_X, typename Dist, typename T>
+    template <int BLOCK_DIM_X, typename T>
     class ReduceDescCalculatorSimple
     {
     public:
@@ -499,29 +373,30 @@ namespace cv { namespace gpu { namespace bfmatcher
             queryDescs = queryDescs_;
         }
 
-        __device__ void calc(const T* trainDescs, int desc_len, float* sdiff_row) const
+        template <typename Dist>
+        __device__ void calc(const T* trainDescs, int desc_len, Dist& dist, float* sdiff_row) const
         {
-            reduceDescDiff<BLOCK_DIM_X, Dist>(queryDescs, trainDescs, desc_len, sdiff_row);
+            reduceDescDiff<BLOCK_DIM_X>(queryDescs, trainDescs, desc_len, dist, sdiff_row);
         }
 
     private:
         const T* queryDescs;
     };
 
-    template <int BLOCK_DIM_X, int BLOCK_DIM_Y, int MAX_DESCRIPTORS_LEN, bool DESC_LEN_EQ_MAX_LEN, 
-        typename Dist, typename T>
-    class ReduceDescCalculatorSmem
+    template <int BLOCK_DIM_X, int MAX_DESCRIPTORS_LEN, bool DESC_LEN_EQ_MAX_LEN, typename T>
+    class ReduceDescCalculatorCached
     {
     public:
         __device__ void prepare(const T* queryDescs, int desc_len, float* smem)
         {
-            loadDescsVals<BLOCK_DIM_X, BLOCK_DIM_Y, MAX_DESCRIPTORS_LEN>(queryDescs, desc_len, smem, queryVals);
+            loadDescsVals<BLOCK_DIM_X, MAX_DESCRIPTORS_LEN>(queryDescs, desc_len, queryVals, smem);
         }
 
-        __device__ void calc(const T* trainDescs, int desc_len, float* sdiff_row) const
+        template <typename Dist>
+        __device__ void calc(const T* trainDescs, int desc_len, Dist& dist, float* sdiff_row) const
         {
-            reduceDescDiff_smem<BLOCK_DIM_X, MAX_DESCRIPTORS_LEN, DESC_LEN_EQ_MAX_LEN, Dist>(queryVals, trainDescs, 
-                desc_len, sdiff_row);
+            reduceDescDiffCached<BLOCK_DIM_X, MAX_DESCRIPTORS_LEN, DESC_LEN_EQ_MAX_LEN>(queryVals, trainDescs, 
+                desc_len, dist, sdiff_row);
         }
 
     private:
@@ -531,26 +406,26 @@ namespace cv { namespace gpu { namespace bfmatcher
     ///////////////////////////////////////////////////////////////////////////////
     // matchDescs loop
 
-    template <typename ReduceDescCalculator, typename T, typename Mask>
-    __device__ void matchDescs(int queryIdx, const int imgIdx, const DevMem2D_<T>& trainDescs_,  
+    template <typename Dist, typename ReduceDescCalculator, typename T, typename Mask>
+    __device__ void matchDescs(int queryIdx, int imgIdx, const DevMem2D_<T>& trainDescs_,  
         const Mask& m, const ReduceDescCalculator& reduceDescCalc,
-        float* sdiff_row, float& myMin, int& myBestTrainIdx, int& myBestImgIdx)
+        float& myMin, int& myBestTrainIdx, int& myBestImgIdx, float* sdiff_row)
     {
-        const T* trainDescs = trainDescs_.ptr(threadIdx.y);
-        const int trainDescsStep = blockDim.y * trainDescs_.step / sizeof(T);
-        for (int trainIdx = threadIdx.y; trainIdx < trainDescs_.rows; 
-             trainIdx += blockDim.y, trainDescs += trainDescsStep)
+        for (int trainIdx = threadIdx.y; trainIdx < trainDescs_.rows; trainIdx += blockDim.y)
         {
             if (m(queryIdx, trainIdx))
             {
-                reduceDescCalc.calc(trainDescs, trainDescs_.cols, sdiff_row);
+                const T* trainDescs = trainDescs_.ptr(trainIdx);
+
+                Dist dist;
+
+                reduceDescCalc.calc(trainDescs, trainDescs_.cols, dist, sdiff_row);
 
                 if (threadIdx.x == 0)
                 {
-                    float reg = sdiff_row[0];
-                    if (reg < myMin)
+                    if (dist < myMin)
                     {
-                        myMin = reg;
+                        myMin = dist;
                         myBestTrainIdx = trainIdx;
                         myBestImgIdx = imgIdx;
                     }
@@ -570,18 +445,19 @@ namespace cv { namespace gpu { namespace bfmatcher
         {
         }
 
-        template <typename ReduceDescCalculator, typename Mask>
+        template <typename Dist, typename ReduceDescCalculator, typename Mask>
         __device__ void loop(int queryIdx, Mask& m, const ReduceDescCalculator& reduceDescCalc, 
-            float* sdiff_row, float& myMin, int& myBestTrainIdx, int& myBestImgIdx) const
+            float& myMin, int& myBestTrainIdx, int& myBestImgIdx, float* sdiff_row) const
         {
-            matchDescs(queryIdx, 0, trainDescs, m, reduceDescCalc, 
-                sdiff_row, myMin, myBestTrainIdx, myBestImgIdx);
+            matchDescs<Dist>(queryIdx, 0, trainDescs, m, reduceDescCalc, 
+                myMin, myBestTrainIdx, myBestImgIdx, sdiff_row);
         }
 
         __device__ int desc_len() const
         {
             return trainDescs.cols;
         }
+
     private:
         DevMem2D_<T> trainDescs;
     };
@@ -595,16 +471,16 @@ namespace cv { namespace gpu { namespace bfmatcher
         {
         }
 
-        template <typename ReduceDescCalculator, typename Mask>
+        template <typename Dist, typename ReduceDescCalculator, typename Mask>
         __device__ void loop(int queryIdx, Mask& m, const ReduceDescCalculator& reduceDescCalc, 
-            float* sdiff_row, float& myMin, int& myBestTrainIdx, int& myBestImgIdx) const
+            float& myMin, int& myBestTrainIdx, int& myBestImgIdx, float* sdiff_row) const
         {
             for (int imgIdx = 0; imgIdx < nImg; ++imgIdx)
             {
                 DevMem2D_<T> trainDescs = trainCollection[imgIdx];
                 m.nextMask();
-                matchDescs(queryIdx, imgIdx, trainDescs, m, reduceDescCalc, 
-                    sdiff_row, myMin, myBestTrainIdx, myBestImgIdx);
+                matchDescs<Dist>(queryIdx, imgIdx, trainDescs, m, reduceDescCalc, 
+                    myMin, myBestTrainIdx, myBestImgIdx, sdiff_row);
             }
         }
 
@@ -612,6 +488,7 @@ namespace cv { namespace gpu { namespace bfmatcher
         {
             return desclen;
         }
+
     private:
         const DevMem2D_<T>* trainCollection;
         int nImg;
@@ -623,12 +500,10 @@ namespace cv { namespace gpu { namespace bfmatcher
 
     template <int BLOCK_DIM_X, int BLOCK_DIM_Y, typename ReduceDescCalculator, typename Dist, typename T, 
         typename Train, typename Mask>
-    __global__ void match(PtrStep_<T> queryDescs_, Train train, Mask mask, int* trainIdx, int* imgIdx, float* distance)
+    __global__ void match(const PtrStep_<T> queryDescs_, const Train train, const Mask mask, 
+        int* trainIdx, int* imgIdx, float* distance)
     {
-        __shared__ float sdiff[BLOCK_DIM_X * BLOCK_DIM_Y];
-        __shared__ float smin[64];
-        __shared__ int strainIdx[64];
-        __shared__ int simgIdx[64];
+        __shared__ float smem[BLOCK_DIM_X * BLOCK_DIM_Y];        
         
         const int queryIdx = blockIdx.x;
         
@@ -637,24 +512,39 @@ namespace cv { namespace gpu { namespace bfmatcher
         float myMin = numeric_limits_gpu<float>::max();
 
         {
-            float* sdiff_row = sdiff + BLOCK_DIM_X * threadIdx.y;
+            float* sdiff_row = smem + BLOCK_DIM_X * threadIdx.y;
+
             Mask m = mask;
+
             ReduceDescCalculator reduceDescCalc;
-            reduceDescCalc.prepare(queryDescs_.ptr(queryIdx), train.desc_len(), sdiff);
+
+            reduceDescCalc.prepare(queryDescs_.ptr(queryIdx), train.desc_len(), smem);
         
-            train.loop(queryIdx, m, reduceDescCalc, sdiff_row, myMin, myBestTrainIdx, myBestImgIdx);
+            train.template loop<Dist>(queryIdx, m, reduceDescCalc, myMin, myBestTrainIdx, myBestImgIdx, sdiff_row);
         }
+        __syncthreads();
+
+        float* smin = smem;
+        int* strainIdx = (int*)(smin + BLOCK_DIM_Y);
+        int* simgIdx = strainIdx + BLOCK_DIM_Y;
 
-        findBestMatch<BLOCK_DIM_Y, Dist>(queryIdx, myMin, myBestTrainIdx, myBestImgIdx, 
-            smin, strainIdx, simgIdx, trainIdx, imgIdx, distance);
+        findBestMatch<BLOCK_DIM_Y>(myMin, myBestTrainIdx, myBestImgIdx, 
+            smin, strainIdx, simgIdx);
+
+        if (threadIdx.x == 0 && threadIdx.y == 0)
+        {
+            imgIdx[queryIdx] = myBestImgIdx;
+            trainIdx[queryIdx] = myBestTrainIdx;
+            distance[queryIdx] = myMin;
+        }
     }
     
     ///////////////////////////////////////////////////////////////////////////////
     // Match kernel callers
 
-    template <int BLOCK_DIM_X, int BLOCK_DIM_Y, template <int> class Dist, typename T, 
+    template <int BLOCK_DIM_X, int BLOCK_DIM_Y, typename Dist, typename T, 
         typename Train, typename Mask>
-    void match_caller(const DevMem2D_<T>& queryDescs, const Train& train, 
+    void matchSimple_caller(const DevMem2D_<T>& queryDescs, const Train& train, 
         const Mask& mask, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance)
     {
         StaticAssert<BLOCK_DIM_Y <= 64>::check(); // blockDimY vals must reduce by warp
@@ -662,15 +552,15 @@ namespace cv { namespace gpu { namespace bfmatcher
         dim3 grid(queryDescs.rows, 1, 1);
         dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y, 1);
 
-        match<BLOCK_DIM_X, BLOCK_DIM_Y, ReduceDescCalculatorSimple<BLOCK_DIM_X, Dist<BLOCK_DIM_X>, T>, 
-            Dist<BLOCK_DIM_X>, T><<<grid, threads>>>(queryDescs, train, mask, trainIdx.data, 
+        match<BLOCK_DIM_X, BLOCK_DIM_Y, ReduceDescCalculatorSimple<BLOCK_DIM_X, T>, Dist, T>
+            <<<grid, threads>>>(queryDescs, train, mask, trainIdx.data, 
             imgIdx.data, distance.data);
 
         cudaSafeCall( cudaThreadSynchronize() );
     }
     template <int BLOCK_DIM_X, int BLOCK_DIM_Y, int MAX_DESCRIPTORS_LEN, bool DESC_LEN_EQ_MAX_LEN, 
-        template <int> class Dist, typename T, typename Train, typename Mask>
-    void match_smem_caller(const DevMem2D_<T>& queryDescs, const Train& train, 
+        typename Dist, typename T, typename Train, typename Mask>
+    void matchCached_caller(const DevMem2D_<T>& queryDescs, const Train& train, 
         const Mask& mask, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance)
     {
         StaticAssert<BLOCK_DIM_Y <= 64>::check();                                // blockDimY vals must reduce by warp
@@ -680,9 +570,10 @@ namespace cv { namespace gpu { namespace bfmatcher
         dim3 grid(queryDescs.rows, 1, 1);
         dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y, 1);
 
-        match<BLOCK_DIM_X, BLOCK_DIM_Y, ReduceDescCalculatorSmem<BLOCK_DIM_X, BLOCK_DIM_Y, 
-              MAX_DESCRIPTORS_LEN, DESC_LEN_EQ_MAX_LEN, Dist<BLOCK_DIM_X>, T>, 
-              Dist<BLOCK_DIM_X>, T><<<grid, threads>>>(queryDescs, train, mask, trainIdx.data, 
+        match<BLOCK_DIM_X, BLOCK_DIM_Y, 
+              ReduceDescCalculatorCached<BLOCK_DIM_X, MAX_DESCRIPTORS_LEN, DESC_LEN_EQ_MAX_LEN, T>, 
+              Dist, T>
+              <<<grid, threads>>>(queryDescs, train, mask, trainIdx.data, 
               imgIdx.data, distance.data);
 
         cudaSafeCall( cudaThreadSynchronize() );
@@ -691,24 +582,24 @@ namespace cv { namespace gpu { namespace bfmatcher
     ///////////////////////////////////////////////////////////////////////////////
     // Match kernel chooser
 
-    template <template <int> class Dist, typename T, typename Train, typename Mask>
+    template <typename Dist, typename T, typename Train, typename Mask>
     void match_chooser(const DevMem2D_<T>& queryDescs, const Train& train, 
         const Mask& mask, const DevMem2Di& trainIdx, const DevMem2Di& imgIdx, const DevMem2Df& distance)
     {
         if (queryDescs.cols < 64)
-            match_smem_caller<16, 16, 64, false, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
+            matchCached_caller<16, 16, 64, false, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
         else if (queryDescs.cols == 64)
-            match_smem_caller<16, 16, 64, true, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
+            matchCached_caller<16, 16, 64, true, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
         else if (queryDescs.cols < 128)
-            match_smem_caller<16, 16, 128, false, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
+            matchCached_caller<16, 16, 128, false, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
         else if (queryDescs.cols == 128)
-            match_smem_caller<16, 16, 128, true, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
+            matchCached_caller<16, 16, 128, true, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
         else if (queryDescs.cols < 256)
-            match_smem_caller<16, 16, 256, false, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
+            matchCached_caller<16, 16, 256, false, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
         else if (queryDescs.cols == 256)
-            match_smem_caller<16, 16, 256, true, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
+            matchCached_caller<16, 16, 256, true, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
         else
-            match_caller<16, 16, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
+            matchSimple_caller<16, 16, Dist>(queryDescs, train, mask, trainIdx, imgIdx, distance);
 
         cudaSafeCall( cudaThreadSynchronize() );
     }
@@ -828,41 +719,41 @@ namespace cv { namespace gpu { namespace bfmatcher
         {
             const T* trainDescs = trainDescs_.ptr(trainIdx);
 
-            float dist = numeric_limits_gpu<float>::max();
+            float myDist = numeric_limits_gpu<float>::max();
 
             if (mask(queryIdx, trainIdx))
             {
-                reduceDescDiff<BLOCK_DIM_X, Dist>(queryDescs, trainDescs, trainDescs_.cols, sdiff_row);
+                Dist dist;
+
+                reduceDescDiff<BLOCK_DIM_X>(queryDescs, trainDescs, trainDescs_.cols, dist, sdiff_row);
 
                 if (threadIdx.x == 0)
-                {
-                    dist = Dist::finalResult(sdiff_row[0]);
-                }
+                    myDist = dist;
             }
             
             if (threadIdx.x == 0)
-                distance.ptr(queryIdx)[trainIdx] = dist;
+                distance.ptr(queryIdx)[trainIdx] = myDist;
         }
     }
 
     ///////////////////////////////////////////////////////////////////////////////
     // Calc distance kernel caller
 
-    template <int BLOCK_DIM_X, int BLOCK_DIM_Y, template <int> class Dist, typename T, typename Mask>
+    template <int BLOCK_DIM_X, int BLOCK_DIM_Y, typename Dist, typename T, typename Mask>
     void calcDistance_caller(const DevMem2D_<T>& queryDescs, const DevMem2D_<T>& trainDescs, 
         const Mask& mask, const DevMem2Df& distance)
     {
         dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y, 1);
         dim3 grid(queryDescs.rows, divUp(trainDescs.rows, BLOCK_DIM_Y), 1);
 
-        calcDistance<BLOCK_DIM_X, BLOCK_DIM_Y, Dist<BLOCK_DIM_X>, T><<<grid, threads>>>(
+        calcDistance<BLOCK_DIM_X, BLOCK_DIM_Y, Dist, T><<<grid, threads>>>(
             queryDescs, trainDescs, mask, distance);
 
         cudaSafeCall( cudaThreadSynchronize() );
     }
         
     ///////////////////////////////////////////////////////////////////////////////
-    // reduceMin
+    // warpReduceMinIdx
 
     template <int BLOCK_SIZE> 
     __device__ void warpReduceMinIdx(volatile float* sdist, volatile int* strainIdx, float& myMin, int tid)
@@ -1103,25 +994,27 @@ namespace cv { namespace gpu { namespace bfmatcher
     {
         #if defined (__CUDA_ARCH__) && __CUDA_ARCH__ >= 110
 
-        __shared__ float sdiff[BLOCK_DIM_X * BLOCK_DIM_Y];
+        __shared__ float smem[BLOCK_DIM_X * BLOCK_DIM_Y];
 
-        float* sdiff_row = sdiff + BLOCK_DIM_X * threadIdx.y;
+        float* sdiff_row = smem + BLOCK_DIM_X * threadIdx.y;
         
         const int queryIdx = blockIdx.x;
         const T* queryDescs = queryDescs_.ptr(queryIdx);
 
         const int trainIdx = blockIdx.y * BLOCK_DIM_Y + threadIdx.y;
+
         if (trainIdx < trainDescs_.rows)
         {
             const T* trainDescs = trainDescs_.ptr(trainIdx);
 
             if (mask(queryIdx, trainIdx))
             {
-                reduceDescDiff<BLOCK_DIM_X, Dist>(queryDescs, trainDescs, trainDescs_.cols, sdiff_row);
+                Dist dist;
+
+                reduceDescDiff<BLOCK_DIM_X>(queryDescs, trainDescs, trainDescs_.cols, dist, sdiff_row);
 
                 if (threadIdx.x == 0)
                 {
-                    float dist = Dist::finalResult(sdiff_row[0]);
                     if (dist < maxDistance)
                     {
                         unsigned int i = atomicInc(nMatches + queryIdx, (unsigned int) -1);
@@ -1141,7 +1034,7 @@ namespace cv { namespace gpu { namespace bfmatcher
     ///////////////////////////////////////////////////////////////////////////////
     // Radius Match kernel caller
 
-    template <int BLOCK_DIM_X, int BLOCK_DIM_Y, template <int> class Dist, typename T, typename Mask>
+    template <int BLOCK_DIM_X, int BLOCK_DIM_Y, typename Dist, typename T, typename Mask>
     void radiusMatch_caller(const DevMem2D_<T>& queryDescs, const DevMem2D_<T>& trainDescs, 
         float maxDistance, const Mask& mask, const DevMem2Di& trainIdx, unsigned int* nMatches, 
         const DevMem2Df& distance)
@@ -1149,7 +1042,7 @@ namespace cv { namespace gpu { namespace bfmatcher
         dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y, 1);
         dim3 grid(queryDescs.rows, divUp(trainDescs.rows, BLOCK_DIM_Y), 1);
 
-        radiusMatch<BLOCK_DIM_X, BLOCK_DIM_Y, Dist<BLOCK_DIM_X>, T><<<grid, threads>>>(
+        radiusMatch<BLOCK_DIM_X, BLOCK_DIM_Y, Dist, T><<<grid, threads>>>(
             queryDescs, trainDescs, maxDistance, mask, trainIdx, nMatches, distance);
 
         cudaSafeCall( cudaThreadSynchronize() );
diff --git a/modules/gpu/src/imgproc_gpu.cpp b/modules/gpu/src/imgproc_gpu.cpp
index 35aa3d5..16d8e6b 100644
--- a/modules/gpu/src/imgproc_gpu.cpp
+++ b/modules/gpu/src/imgproc_gpu.cpp
@@ -66,7 +66,10 @@ void cv::gpu::integral(const GpuMat&, GpuMat&, GpuMat&) { throw_nogpu(); }
 void cv::gpu::sqrIntegral(const GpuMat&, GpuMat&) { throw_nogpu(); }
 void cv::gpu::columnSum(const GpuMat&, GpuMat&) { throw_nogpu(); }
 void cv::gpu::rectStdDev(const GpuMat&, const GpuMat&, GpuMat&, const Rect&) { throw_nogpu(); }
-void cv::gpu::Canny(const GpuMat&, GpuMat&, double, double, int) { throw_nogpu(); }
+//void cv::gpu::Canny(const GpuMat&, GpuMat&, double, double, int) { throw_nogpu(); }
+//void cv::gpu::Canny(const GpuMat&, GpuMat&, GpuMat&, double, double, int) { throw_nogpu(); }
+//void cv::gpu::Canny(const GpuMat&, const GpuMat&, GpuMat&, GpuMat&, double, double, int) { throw_nogpu(); }
+//void cv::gpu::Canny(const GpuMat&, const GpuMat&, GpuMat&, GpuMat&, GpuMat&, double, double, int) { throw_nogpu(); }
 void cv::gpu::evenLevels(GpuMat&, int, int, int) { throw_nogpu(); }
 void cv::gpu::histEven(const GpuMat&, GpuMat&, int, int, int) { throw_nogpu(); }
 void cv::gpu::histEven(const GpuMat&, GpuMat*, int*, int*, int*) { throw_nogpu(); }
@@ -655,34 +658,60 @@ void cv::gpu::rectStdDev(const GpuMat& src, const GpuMat& sqr, GpuMat& dst, cons
 ////////////////////////////////////////////////////////////////////////
 // Canny
 
-void cv::gpu::Canny(const GpuMat& image, GpuMat& edges, double threshold1, double threshold2, int apertureSize)
-{
-    CV_Assert(!"disabled until fix crash");
-    CV_Assert(image.type() == CV_8UC1);
-
-    GpuMat srcDx, srcDy;
-
-    Sobel(image, srcDx, -1, 1, 0, apertureSize);
-    Sobel(image, srcDy, -1, 0, 1, apertureSize);
-
-    srcDx.convertTo(srcDx, CV_32F);
-    srcDy.convertTo(srcDy, CV_32F);
-
-    edges.create(image.size(), CV_8UC1);
-
-    NppiSize sz;
-    sz.height = image.rows;
-    sz.width = image.cols;
-
-    int bufsz;
-    nppSafeCall( nppiCannyGetBufferSize(sz, &bufsz) );
-    GpuMat buf(1, bufsz, CV_8UC1);
-
-    nppSafeCall( nppiCanny_32f8u_C1R(srcDx.ptr<Npp32f>(), srcDx.step, srcDy.ptr<Npp32f>(), srcDy.step,
-        edges.ptr<Npp8u>(), edges.step, sz, (Npp32f)threshold1, (Npp32f)threshold2, buf.ptr<Npp8u>()) );
-
-    cudaSafeCall( cudaThreadSynchronize() );
-}
+//void cv::gpu::Canny(const GpuMat& image, GpuMat& edges, double threshold1, double threshold2, int apertureSize)
+//{
+//    CV_Assert(!"disabled until fix crash");
+//
+//    GpuMat srcDx, srcDy;
+//
+//    Sobel(image, srcDx, CV_32F, 1, 0, apertureSize);
+//    Sobel(image, srcDy, CV_32F, 0, 1, apertureSize);
+//
+//    GpuMat buf;
+//
+//    Canny(srcDx, srcDy, edges, buf, threshold1, threshold2, apertureSize);
+//}
+//
+//void cv::gpu::Canny(const GpuMat& image, GpuMat& edges, GpuMat& buf, double threshold1, double threshold2, int apertureSize)
+//{
+//    CV_Assert(!"disabled until fix crash");
+//
+//    GpuMat srcDx, srcDy;
+//
+//    Sobel(image, srcDx, CV_32F, 1, 0, apertureSize);
+//    Sobel(image, srcDy, CV_32F, 0, 1, apertureSize);
+//
+//    Canny(srcDx, srcDy, edges, buf, threshold1, threshold2, apertureSize);
+//}
+//
+//void cv::gpu::Canny(const GpuMat& srcDx, const GpuMat& srcDy, GpuMat& edges, double threshold1, double threshold2, int apertureSize)
+//{
+//    CV_Assert(!"disabled until fix crash");
+//
+//    GpuMat buf;
+//    Canny(srcDx, srcDy, edges, buf, threshold1, threshold2, apertureSize);
+//}
+//
+//void cv::gpu::Canny(const GpuMat& srcDx, const GpuMat& srcDy, GpuMat& edges, GpuMat& buf, double threshold1, double threshold2, int apertureSize)
+//{
+//    CV_Assert(!"disabled until fix crash");
+//    CV_Assert(srcDx.type() == CV_32FC1 && srcDy.type() == CV_32FC1 && srcDx.size() == srcDy.size());
+//
+//    edges.create(srcDx.size(), CV_8UC1);
+//
+//    NppiSize sz;
+//    sz.height = srcDx.rows;
+//    sz.width = srcDx.cols;
+//
+//    int bufsz;
+//    nppSafeCall( nppiCannyGetBufferSize(sz, &bufsz) );
+//    ensureSizeIsEnough(1, bufsz, CV_8UC1, buf);
+//
+//    nppSafeCall( nppiCanny_32f8u_C1R(srcDx.ptr<Npp32f>(), srcDx.step, srcDy.ptr<Npp32f>(), srcDy.step,
+//        edges.ptr<Npp8u>(), edges.step, sz, (Npp32f)threshold1, (Npp32f)threshold2, buf.ptr<Npp8u>()) );
+//
+//    cudaSafeCall( cudaThreadSynchronize() );
+//}
 
 ////////////////////////////////////////////////////////////////////////
 // Histogram
diff --git a/tests/gpu/src/arithm.cpp b/tests/gpu/src/arithm.cpp
index fe2385a..d03642a 100644
--- a/tests/gpu/src/arithm.cpp
+++ b/tests/gpu/src/arithm.cpp
@@ -66,45 +66,58 @@ protected:
 
     virtual int test(const Mat& mat1, const Mat& mat2) = 0;
 
-    int CheckNorm(const Mat& m1, const Mat& m2);
-    int CheckNorm(const Scalar& s1, const Scalar& s2);
-    int CheckNorm(double d1, double d2);
+    int CheckNorm(const Mat& m1, const Mat& m2, double eps = 1e-5);
+    int CheckNorm(const Scalar& s1, const Scalar& s2, double eps = 1e-5);
+    int CheckNorm(double d1, double d2, double eps = 1e-5);
 };
 
 int CV_GpuArithmTest::test(int type)
 {
     cv::Size sz(200, 200);
     cv::Mat mat1(sz, type), mat2(sz, type);
+    
     cv::RNG rng(*ts->get_rng());
-    rng.fill(mat1, cv::RNG::UNIFORM, cv::Scalar::all(1), cv::Scalar::all(20));
-    rng.fill(mat2, cv::RNG::UNIFORM, cv::Scalar::all(1), cv::Scalar::all(20));
+
+    if (type != CV_32FC1)
+    {
+        rng.fill(mat1, cv::RNG::UNIFORM, cv::Scalar::all(1), cv::Scalar::all(20));
+        rng.fill(mat2, cv::RNG::UNIFORM, cv::Scalar::all(1), cv::Scalar::all(20));
+    }
+    else
+    {
+        rng.fill(mat1, cv::RNG::UNIFORM, cv::Scalar::all(0.1), cv::Scalar::all(1.0));
+        rng.fill(mat2, cv::RNG::UNIFORM, cv::Scalar::all(0.1), cv::Scalar::all(1.0));
+    }
 
     return test(mat1, mat2);
 }
 
-int CV_GpuArithmTest::CheckNorm(const Mat& m1, const Mat& m2)
+int CV_GpuArithmTest::CheckNorm(const Mat& m1, const Mat& m2, double eps)
 {
     double ret = norm(m1, m2, NORM_INF);
 
-    if (ret < 1e-5)
+    if (ret < eps)
         return CvTS::OK;
 
     ts->printf(CvTS::LOG, "\nNorm: %f\n", ret);
     return CvTS::FAIL_GENERIC;
 }
 
-int CV_GpuArithmTest::CheckNorm(const Scalar& s1, const Scalar& s2)
+int CV_GpuArithmTest::CheckNorm(const Scalar& s1, const Scalar& s2, double eps)
 {
-    double ret0 = CheckNorm(s1[0], s2[0]), ret1 = CheckNorm(s1[1], s2[1]), ret2 = CheckNorm(s1[2], s2[2]), ret3 = CheckNorm(s1[3], s2[3]);
+    int ret0 = CheckNorm(s1[0], s2[0], eps), 
+        ret1 = CheckNorm(s1[1], s2[1], eps), 
+        ret2 = CheckNorm(s1[2], s2[2], eps), 
+        ret3 = CheckNorm(s1[3], s2[3], eps);
 
     return (ret0 == CvTS::OK && ret1 == CvTS::OK && ret2 == CvTS::OK && ret3 == CvTS::OK) ? CvTS::OK : CvTS::FAIL_GENERIC;
 }
 
-int CV_GpuArithmTest::CheckNorm(double d1, double d2)
+int CV_GpuArithmTest::CheckNorm(double d1, double d2, double eps)
 {
     double ret = ::fabs(d1 - d2);
 
-    if (ret < 1e-5)
+    if (ret < eps)
         return CvTS::OK;
 
     ts->printf(CvTS::LOG, "\nNorm: %f\n", ret);
@@ -245,7 +258,7 @@ struct CV_GpuNppImageDivideTest : public CV_GpuArithmTest
 	    GpuMat gpuRes;
 	    cv::gpu::divide(gpu1, gpu2, gpuRes);
 
-            return CheckNorm(cpuRes, gpuRes);
+        return CheckNorm(cpuRes, gpuRes, 1.01f);
     }
 };
 
@@ -584,7 +597,7 @@ struct CV_GpuNppImagePhaseTest : public CV_GpuArithmTest
         GpuMat gpuRes;
         cv::gpu::phase(gpu1, gpu2, gpuRes, true);
 
-        return CheckNorm(cpuRes, gpuRes);
+        return CheckNorm(cpuRes, gpuRes, 0.3f);
     }
 };
 
@@ -611,7 +624,7 @@ struct CV_GpuNppImageCartToPolarTest : public CV_GpuArithmTest
         cv::gpu::cartToPolar(gpu1, gpu2, gpuMag, gpuAngle);
 
         int magRes = CheckNorm(cpuMag, gpuMag);
-        int angleRes = CheckNorm(cpuAngle, gpuAngle);
+        int angleRes = CheckNorm(cpuAngle, gpuAngle, 0.005f);
 
         return magRes == CvTS::OK && angleRes == CvTS::OK ? CvTS::OK : CvTS::FAIL_GENERIC;
     }
diff --git a/tests/gpu/src/brute_force_matcher.cpp b/tests/gpu/src/brute_force_matcher.cpp
index afe5db2..f4c0acc 100644
--- a/tests/gpu/src/brute_force_matcher.cpp
+++ b/tests/gpu/src/brute_force_matcher.cpp
@@ -51,24 +51,27 @@ class CV_GpuBruteForceMatcherTest : public CvTest
 {
 public:
     CV_GpuBruteForceMatcherTest() :
-        CvTest( "GPU-BruteForceMatcher", "BruteForceMatcher" ), badPart(0.01f)
+        CvTest( "GPU-BruteForceMatcher", "BruteForceMatcher" )
     {
     }
-protected:
-    static const int dim = 500;
-    static const int queryDescCount = 300; // must be even number because we split train data in some cases in two
-    static const int countFactor = 4; // do not change it
-    const float badPart;
 
+protected:
     virtual void run(int);
-    void generateData(GpuMat& query, GpuMat& train);
-
+    
     void emptyDataTest();
+    void dataTest(int dim);
+    
+    void generateData(GpuMat& query, GpuMat& train, int dim);
+
     void matchTest(const GpuMat& query, const GpuMat& train);
     void knnMatchTest(const GpuMat& query, const GpuMat& train);
     void radiusMatchTest(const GpuMat& query, const GpuMat& train);
 
+private:
     BruteForceMatcher_GPU< L2<float> > dmatcher;
+
+    static const int queryDescCount = 300; // must be even number because we split train data in some cases in two
+    static const int countFactor = 4; // do not change it
 };
 
 void CV_GpuBruteForceMatcherTest::emptyDataTest()
@@ -150,7 +153,7 @@ void CV_GpuBruteForceMatcherTest::emptyDataTest()
 
 }
 
-void CV_GpuBruteForceMatcherTest::generateData( GpuMat& queryGPU, GpuMat& trainGPU )
+void CV_GpuBruteForceMatcherTest::generateData( GpuMat& queryGPU, GpuMat& trainGPU, int dim )
 {
     Mat query, train;
     RNG rng(*ts->get_rng());
@@ -209,7 +212,7 @@ void CV_GpuBruteForceMatcherTest::matchTest( const GpuMat& query, const GpuMat&
                 if( (match.queryIdx != (int)i) || (match.trainIdx != (int)i*countFactor) || (match.imgIdx != 0) )
                     badCount++;
             }
-            if( (float)badCount > (float)queryDescCount*badPart )
+            if (badCount > 0)
             {
                 ts->printf( CvTS::LOG, "%f - too large bad matches part while test match() function (1).\n",
                             (float)badCount/(float)queryDescCount );
@@ -260,7 +263,7 @@ void CV_GpuBruteForceMatcherTest::matchTest( const GpuMat& query, const GpuMat&
                     }
                 }
             }
-            if( (float)badCount > (float)queryDescCount*badPart )
+            if (badCount > 0)
             {
                 ts->printf( CvTS::LOG, "%f - too large bad matches part while test match() function (2).\n",
                             (float)badCount/(float)queryDescCount );
@@ -305,7 +308,7 @@ void CV_GpuBruteForceMatcherTest::knnMatchTest( const GpuMat& query, const GpuMa
                     badCount += localBadCount > 0 ? 1 : 0;
                 }
             }
-            if( (float)badCount > (float)queryDescCount*badPart )
+            if (badCount > 0)
             {
                 ts->printf( CvTS::LOG, "%f - too large bad matches part while test knnMatch() function (1).\n",
                             (float)badCount/(float)queryDescCount );
@@ -369,7 +372,7 @@ void CV_GpuBruteForceMatcherTest::knnMatchTest( const GpuMat& query, const GpuMa
                     badCount += localBadCount > 0 ? 1 : 0;
                 }
             }
-            if( (float)badCount > (float)queryDescCount*badPart )
+            if (badCount > 0)
             {
                 ts->printf( CvTS::LOG, "%f - too large bad matches part while test knnMatch() function (2).\n",
                             (float)badCount/(float)queryDescCount );
@@ -407,7 +410,7 @@ void CV_GpuBruteForceMatcherTest::radiusMatchTest( const GpuMat& query, const Gp
                         badCount++;
                 }
             }
-            if( (float)badCount > (float)queryDescCount*badPart )
+            if (badCount > 0)
             {
                 ts->printf( CvTS::LOG, "%f - too large bad matches part while test radiusMatch() function (1).\n",
                             (float)badCount/(float)queryDescCount );
@@ -473,7 +476,8 @@ void CV_GpuBruteForceMatcherTest::radiusMatchTest( const GpuMat& query, const Gp
                 badCount += localBadCount > 0 ? 1 : 0;
             }
         }
-        if( (float)badCount > (float)queryDescCount*badPart )
+
+        if (badCount > 0)
         {
             curRes = CvTS::FAIL_INVALID_OUTPUT;
             ts->printf( CvTS::LOG, "%f - too large bad matches part while test radiusMatch() function (2).\n",
@@ -483,20 +487,29 @@ void CV_GpuBruteForceMatcherTest::radiusMatchTest( const GpuMat& query, const Gp
     }
 }
 
-void CV_GpuBruteForceMatcherTest::run( int )
+void CV_GpuBruteForceMatcherTest::dataTest(int dim)
 {
-    emptyDataTest();
-
     GpuMat query, train;
-    generateData( query, train );
+    generateData(query, train, dim);
 
-    matchTest( query, train );
+    matchTest(query, train);
+    knnMatchTest(query, train);
+    radiusMatchTest(query, train);
 
-    knnMatchTest( query, train );
+    dmatcher.clear();
+}
 
-    radiusMatchTest( query, train );
+void CV_GpuBruteForceMatcherTest::run(int)
+{
+    emptyDataTest();
 
-    dmatcher.clear();
+    dataTest(50);
+    dataTest(64);
+    dataTest(100);
+    dataTest(128);
+    dataTest(200);
+    dataTest(256);
+    dataTest(300);
 }
 
 CV_GpuBruteForceMatcherTest CV_GpuBruteForceMatcher_test;
diff --git a/tests/gpu/src/gputest_main.cpp b/tests/gpu/src/gputest_main.cpp
index cbeb0d0..5d07d34 100644
--- a/tests/gpu/src/gputest_main.cpp
+++ b/tests/gpu/src/gputest_main.cpp
@@ -45,7 +45,6 @@ CvTS test_system("gpu");
 
 const char* blacklist[] =
 {
-    "GPU-AsyncGpuMatOperator",     // crash
     "GPU-NppImageCanny",            // NPP_TEXTURE_BIND_ERROR
     0
 };
diff --git a/tests/gpu/src/imgproc_gpu.cpp b/tests/gpu/src/imgproc_gpu.cpp
index d3affcf..d941a00 100644
--- a/tests/gpu/src/imgproc_gpu.cpp
+++ b/tests/gpu/src/imgproc_gpu.cpp
@@ -408,30 +408,30 @@ struct CV_GpuNppImageIntegralTest : public CV_GpuImageProcTest
 
 ////////////////////////////////////////////////////////////////////////////////
 // Canny
-struct CV_GpuNppImageCannyTest : public CV_GpuImageProcTest
-{
-    CV_GpuNppImageCannyTest() : CV_GpuImageProcTest( "GPU-NppImageCanny", "Canny" ) {}
-
-    int test(const Mat& img)
-    {
-        if (img.type() != CV_8UC1)
-        {
-            ts->printf(CvTS::LOG, "\nUnsupported type\n");
-            return CvTS::OK;
-        }
-
-        const double threshold1 = 1.0, threshold2 = 10.0;
-
-        Mat cpudst;
-        cv::Canny(img, cpudst, threshold1, threshold2);
-
-        GpuMat gpu1(img);
-        GpuMat gpudst;
-        cv::gpu::Canny(gpu1, gpudst, threshold1, threshold2);
-
-        return CheckNorm(cpudst, gpudst);
-    }
-};
+//struct CV_GpuNppImageCannyTest : public CV_GpuImageProcTest
+//{
+//    CV_GpuNppImageCannyTest() : CV_GpuImageProcTest( "GPU-NppImageCanny", "Canny" ) {}
+//
+//    int test(const Mat& img)
+//    {
+//        if (img.type() != CV_8UC1)
+//        {
+//            ts->printf(CvTS::LOG, "\nUnsupported type\n");
+//            return CvTS::OK;
+//        }
+//
+//        const double threshold1 = 1.0, threshold2 = 10.0;
+//
+//        Mat cpudst;
+//        cv::Canny(img, cpudst, threshold1, threshold2);
+//
+//        GpuMat gpu1(img);
+//        GpuMat gpudst;
+//        cv::gpu::Canny(gpu1, gpudst, threshold1, threshold2);
+//
+//        return CheckNorm(cpudst, gpudst);
+//    }
+//};
 
 ////////////////////////////////////////////////////////////////////////////////
 // cvtColor
@@ -839,7 +839,7 @@ CV_GpuNppImageCopyMakeBorderTest CV_GpuNppImageCopyMakeBorder_test;
 CV_GpuNppImageWarpAffineTest CV_GpuNppImageWarpAffine_test;
 CV_GpuNppImageWarpPerspectiveTest CV_GpuNppImageWarpPerspective_test;
 CV_GpuNppImageIntegralTest CV_GpuNppImageIntegral_test;
-CV_GpuNppImageCannyTest CV_GpuNppImageCanny_test;
+//CV_GpuNppImageCannyTest CV_GpuNppImageCanny_test;
 CV_GpuCvtColorTest CV_GpuCvtColor_test;
 CV_GpuHistogramsTest CV_GpuHistograms_test;
 CV_GpuCornerHarrisTest CV_GpuCornerHarris_test;
diff --git a/tests/gpu/src/operator_async_call.cpp b/tests/gpu/src/operator_async_call.cpp
index 4681327..62fa120 100644
--- a/tests/gpu/src/operator_async_call.cpp
+++ b/tests/gpu/src/operator_async_call.cpp
@@ -40,119 +40,54 @@
 //M*/
 
 #include "gputest.hpp"
-#include <string>
-#include <iostream>
-#include <fstream>
-#include <iterator>
-#include <limits>
-#include <numeric>
-#include <iomanip> // for  cout << setw()
 
-using namespace cv;
 using namespace std;
-using namespace gpu;
-
-class CV_AsyncGpuMatTest : public CvTest
-{
-    public:
-        CV_AsyncGpuMatTest() : CvTest( "GPU-AsyncGpuMatOperator", "async" )
-        {
-             rows = 234;
-            cols = 123;
-
-        }
-        ~CV_AsyncGpuMatTest() {}
-
-    protected:
-        void run(int);
-        template <typename T>
-        void print_mat(const T & mat, const std::string & name) const;
-        bool compare_matrix(cv::Mat & cpumat);
-
-    private:
-        int rows;
-        int cols;
-};
-
-template<typename T>
-void CV_AsyncGpuMatTest::print_mat(const T & mat, const std::string & name) const { cv::imshow(name, mat); }
+using namespace cv;
+using namespace cv::gpu;
 
-bool CV_AsyncGpuMatTest::compare_matrix(cv::Mat & cpumat)
+struct CV_AsyncGpuMatTest : public CvTest
 {
-    Mat cmat(cpumat.size(), cpumat.type(), Scalar::all(0));
-    GpuMat gmat0(cmat);
-    GpuMat gmat1;
-    GpuMat gmat2;
-    GpuMat gmat3;
-
-    //int64 time = getTickCount();
-
-    Stream stream;
-	stream.enqueueMemSet(gmat0, cv::Scalar::all(1), gmat1);
-	stream.enqueueMemSet(gmat0, cv::Scalar::all(1), gmat2);
-	stream.enqueueMemSet(gmat0, cv::Scalar::all(1), gmat3);
-    stream.waitForCompletion();
-
-    //int64 time1 = getTickCount();
-
-    gmat1.copyTo(gmat0);
-    gmat2.copyTo(gmat0);
-    gmat3.copyTo(gmat0);
-
-    //int64 time2 = getTickCount();
-
-    //std::cout << "\ntime async: " << std::fixed << std::setprecision(12) << double((time1 - time)  / (double)getTickFrequency());
-    //std::cout << "\ntime  sync: " << std::fixed << std::setprecision(12) << double((time2 - time1) / (double)getTickFrequency());
-    //std::cout << "\n";
-
-#ifdef PRINT_MATRIX
-    print_mat(cmat, "cpu mat");
-    print_mat(gmat0, "gpu mat 0");
-    print_mat(gmat1, "gpu mat 1");
-    print_mat(gmat2, "gpu mat 2");
-    print_mat(gmat3, "gpu mat 3");
-    cv::waitKey(0);
-#endif
-
-    double ret = norm(cmat, gmat0) + norm(cmat, gmat1) + norm(cmat, gmat2) + norm(cmat, gmat3);
-
-    if (ret < 1.0)
-        return true;
-    else
+    CV_AsyncGpuMatTest() : CvTest( "GPU-AsyncGpuMatOperator", "async" )
     {
-        ts->printf(CvTS::LOG, "\nNorm: %f\n", ret);
-        return false;
     }
-}
 
-void CV_AsyncGpuMatTest::run( int /* start_from */)
-{
-    bool is_test_good = true;
+    void run(int)
+    {
+        try
+        {
+            CudaMem src(Mat::zeros(100, 100, CV_8UC1));
 
-    Mat cpumat(rows, cols, CV_8U);
-    cpumat.setTo(Scalar::all(127));
+            GpuMat gpusrc;
+            GpuMat gpudst0, gpudst1(100, 100, CV_8UC1);
 
-    try
-    {
-        is_test_good &= compare_matrix(cpumat);
-    }
-    catch(cv::Exception& e)
-    {
-        if (!check_and_treat_gpu_exception(e, ts))
-            throw; 
-        return;
-    }
+            CudaMem cpudst0;
+            CudaMem cpudst1;
 
-    if (is_test_good == true)
-        ts->set_failed_test_info(CvTS::OK);
-    else
-        ts->set_failed_test_info(CvTS::FAIL_GENERIC);
-}
+            Stream stream0, stream1;
 
+            stream0.enqueueUpload(src, gpusrc);
+            bitwise_not(gpusrc, gpudst0, GpuMat(), stream0);
+            stream0.enqueueDownload(gpudst0, cpudst0);
 
-/////////////////////////////////////////////////////////////////////////////
-/////////////////// tests registration  /////////////////////////////////////
-/////////////////////////////////////////////////////////////////////////////
+            stream1.enqueueMemSet(gpudst1, Scalar::all(128));
+            stream1.enqueueDownload(gpudst1, cpudst1);
+            
+            stream0.waitForCompletion();
+            stream1.waitForCompletion();
 
+            Mat cpu_gold0(100, 100, CV_8UC1, Scalar::all(255));
+            Mat cpu_gold1(100, 100, CV_8UC1, Scalar::all(128));
 
-CV_AsyncGpuMatTest CV_AsyncGpuMatTest_test;
+            if (norm(cpudst0, cpu_gold0, NORM_INF) > 0 || norm(cpudst1, cpu_gold1, NORM_INF) > 0)
+                ts->set_failed_test_info(CvTS::FAIL_GENERIC);
+            else
+                ts->set_failed_test_info(CvTS::OK);
+        }
+        catch(cv::Exception& e)
+        {
+            if (!check_and_treat_gpu_exception(e, ts))
+                throw; 
+            return;
+        }
+    }
+} CV_AsyncGpuMatTest_test;