From 21c4753fedecdf674058e8d69d3184a2f8c5a021 Mon Sep 17 00:00:00 2001
From: Vladislav Vinogradov <vlad.vinogradov@itseez.com>
Date: Fri, 23 Aug 2013 13:46:29 +0400
Subject: [PATCH] used global memory access for up-scaling

---
 modules/gpu/src/cuda/resize.cu | 677 ++++++++++++++++++++++-------------------
 modules/gpu/src/resize.cpp     |  26 +-
 2 files changed, 380 insertions(+), 323 deletions(-)

diff --git a/modules/gpu/src/cuda/resize.cu b/modules/gpu/src/cuda/resize.cu
index 6ecb7eb..dc9f462 100644
--- a/modules/gpu/src/cuda/resize.cu
+++ b/modules/gpu/src/cuda/resize.cu
@@ -52,368 +52,431 @@
 
 namespace cv { namespace gpu { namespace device
 {
-    namespace imgproc
+    // kernels
+
+    template <typename T> __global__ void resize_nearest(const PtrStep<T> src, PtrStepSz<T> dst, const float fy, const float fx)
     {
-        template <typename T> __global__ void resize_nearest(const PtrStep<T> src, const float fx, const float fy, PtrStepSz<T> dst)
-        {
-            const int dst_x = blockDim.x * blockIdx.x + threadIdx.x;
-            const int dst_y = blockDim.y * blockIdx.y + threadIdx.y;
+        const int dst_x = blockDim.x * blockIdx.x + threadIdx.x;
+        const int dst_y = blockDim.y * blockIdx.y + threadIdx.y;
 
-            if (dst_x < dst.cols && dst_y < dst.rows)
-            {
-                const float src_x = dst_x * fx;
-                const float src_y = dst_y * fy;
+        if (dst_x < dst.cols && dst_y < dst.rows)
+        {
+            const float src_x = dst_x * fx;
+            const float src_y = dst_y * fy;
 
-                dst(dst_y, dst_x) = src(__float2int_rz(src_y), __float2int_rz(src_x));
-            }
+            dst(dst_y, dst_x) = src(__float2int_rz(src_y), __float2int_rz(src_x));
         }
+    }
 
-        template <typename T> __global__ void resize_linear(const PtrStepSz<T> src, const float fx, const float fy, PtrStepSz<T> dst)
-        {
-            typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
+    template <typename T> __global__ void resize_linear(const PtrStepSz<T> src, PtrStepSz<T> dst, const float fy, const float fx)
+    {
+        typedef typename TypeVec<float, VecTraits<T>::cn>::vec_type work_type;
 
-            const int dst_x = blockDim.x * blockIdx.x + threadIdx.x;
-            const int dst_y = blockDim.y * blockIdx.y + threadIdx.y;
+        const int dst_x = blockDim.x * blockIdx.x + threadIdx.x;
+        const int dst_y = blockDim.y * blockIdx.y + threadIdx.y;
 
-            if (dst_x < dst.cols && dst_y < dst.rows)
-            {
-                const float src_x = dst_x * fx;
-                const float src_y = dst_y * fy;
+        if (dst_x < dst.cols && dst_y < dst.rows)
+        {
+            const float src_x = dst_x * fx;
+            const float src_y = dst_y * fy;
 
-                work_type out = VecTraits<work_type>::all(0);
+            work_type out = VecTraits<work_type>::all(0);
 
-                const int x1 = __float2int_rd(src_x);
-                const int y1 = __float2int_rd(src_y);
-                const int x2 = x1 + 1;
-                const int y2 = y1 + 1;
-                const int x2_read = ::min(x2, src.cols - 1);
-                const int y2_read = ::min(y2, src.rows - 1);
+            const int x1 = __float2int_rd(src_x);
+            const int y1 = __float2int_rd(src_y);
+            const int x2 = x1 + 1;
+            const int y2 = y1 + 1;
+            const int x2_read = ::min(x2, src.cols - 1);
+            const int y2_read = ::min(y2, src.rows - 1);
 
-                T src_reg = src(y1, x1);
-                out = out + src_reg * ((x2 - src_x) * (y2 - src_y));
+            T src_reg = src(y1, x1);
+            out = out + src_reg * ((x2 - src_x) * (y2 - src_y));
 
-                src_reg = src(y1, x2_read);
-                out = out + src_reg * ((src_x - x1) * (y2 - src_y));
+            src_reg = src(y1, x2_read);
+            out = out + src_reg * ((src_x - x1) * (y2 - src_y));
 
-                src_reg = src(y2_read, x1);
-                out = out + src_reg * ((x2 - src_x) * (src_y - y1));
+            src_reg = src(y2_read, x1);
+            out = out + src_reg * ((x2 - src_x) * (src_y - y1));
 
-                src_reg = src(y2_read, x2_read);
-                out = out + src_reg * ((src_x - x1) * (src_y - y1));
+            src_reg = src(y2_read, x2_read);
+            out = out + src_reg * ((src_x - x1) * (src_y - y1));
 
-                dst(dst_y, dst_x) = saturate_cast<T>(out);
-            }
+            dst(dst_y, dst_x) = saturate_cast<T>(out);
         }
+    }
 
-        template <class Ptr2D, typename T> __global__ void resize(const Ptr2D src, const float fx, const float fy, PtrStepSz<T> dst)
-        {
-            const int dst_x = blockDim.x * blockIdx.x + threadIdx.x;
-            const int dst_y = blockDim.y * blockIdx.y + threadIdx.y;
+    template <class Ptr2D, typename T> __global__ void resize(const Ptr2D src, PtrStepSz<T> dst, const float fy, const float fx)
+    {
+        const int dst_x = blockDim.x * blockIdx.x + threadIdx.x;
+        const int dst_y = blockDim.y * blockIdx.y + threadIdx.y;
 
-            if (dst_x < dst.cols && dst_y < dst.rows)
-            {
-                const float src_x = dst_x * fx;
-                const float src_y = dst_y * fy;
+        if (dst_x < dst.cols && dst_y < dst.rows)
+        {
+            const float src_x = dst_x * fx;
+            const float src_y = dst_y * fy;
 
-                dst(dst_y, dst_x) = src(src_y, src_x);
-            }
+            dst(dst_y, dst_x) = src(src_y, src_x);
         }
+    }
 
-        template <template <typename> class Filter, typename T> struct ResizeDispatcherStream
+    template <typename Ptr2D, typename T> __global__ void resize_area(const Ptr2D src, PtrStepSz<T> dst)
+    {
+        const int x = blockDim.x * blockIdx.x + threadIdx.x;
+        const int y = blockDim.y * blockIdx.y + threadIdx.y;
+
+        if (x < dst.cols && y < dst.rows)
         {
-            static void call(PtrStepSz<T> src, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
-            {
-                const dim3 block(32, 8);
-                const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+            dst(y, x) = src(y, x);
+        }
+    }
 
-                BrdReplicate<T> brd(src.rows, src.cols);
-                BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
-                Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filteredSrc(brdSrc);
+    // textures
 
-                resize<<<grid, block, 0, stream>>>(filteredSrc, fx, fy, dst);
-                cudaSafeCall( cudaGetLastError() );
-            }
+    template <typename T> struct TextureAccessor;
+
+    #define OPENCV_GPU_IMPLEMENT_RESIZE_TEX(type) \
+        texture<type, cudaTextureType2D, cudaReadModeElementType> tex_resize_##type (0, cudaFilterModePoint, cudaAddressModeClamp); \
+        template <> struct TextureAccessor<type> \
+        { \
+            typedef type elem_type; \
+            typedef int index_type; \
+            int xoff; \
+            int yoff; \
+            __device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
+            { \
+                return tex2D(tex_resize_##type, x + xoff, y + yoff); \
+            } \
+            __host__ static void bind(const PtrStepSz<type>& mat) \
+            { \
+                bindTexture(&tex_resize_##type, mat); \
+            } \
         };
-        template <typename T> struct ResizeDispatcherStream<PointFilter, T>
+
+    OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar)
+    OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar4)
+
+    OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort)
+    OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort4)
+
+    OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short)
+    OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short4)
+
+    OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float)
+    OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float4)
+
+    #undef OPENCV_GPU_IMPLEMENT_RESIZE_TEX
+
+    template <typename T>
+    TextureAccessor<T> texAccessor(const PtrStepSz<T>& mat, int yoff, int xoff)
+    {
+        TextureAccessor<T>::bind(mat);
+
+        TextureAccessor<T> t;
+        t.xoff = xoff;
+        t.yoff = yoff;
+
+        return t;
+    }
+
+    // callers for nearest interpolation
+
+    template <typename T>
+    void call_resize_nearest_glob(const PtrStepSz<T>& src, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
+    {
+        const dim3 block(32, 8);
+        const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+
+        resize_nearest<<<grid, block, 0, stream>>>(src, dst, fy, fx);
+        cudaSafeCall( cudaGetLastError() );
+
+        if (stream == 0)
+            cudaSafeCall( cudaDeviceSynchronize() );
+    }
+
+    template <typename T>
+    void call_resize_nearest_tex(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx)
+    {
+        const dim3 block(32, 8);
+        const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+
+        resize<<<grid, block>>>(texAccessor(srcWhole, yoff, xoff), dst, fy, fx);
+        cudaSafeCall( cudaGetLastError() );
+
+        cudaSafeCall( cudaDeviceSynchronize() );
+    }
+
+    // callers for linear interpolation
+
+    template <typename T>
+    void call_resize_linear_glob(const PtrStepSz<T>& src, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
+    {
+        const dim3 block(32, 8);
+        const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+
+        resize_linear<<<grid, block>>>(src, dst, fy, fx);
+        cudaSafeCall( cudaGetLastError() );
+
+        if (stream == 0)
+            cudaSafeCall( cudaDeviceSynchronize() );
+    }
+
+    template <typename T>
+    void call_resize_linear_tex(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx)
+    {
+        const dim3 block(32, 8);
+        const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+
+        if (srcWhole.data == src.data)
         {
-            static void call(PtrStepSz<T> src, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
-            {
-                const dim3 block(32, 8);
-                const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+            TextureAccessor<T> texSrc = texAccessor(src, 0, 0);
+            LinearFilter< TextureAccessor<T> > filteredSrc(texSrc);
 
-                resize_nearest<<<grid, block, 0, stream>>>(src, fx, fy, dst);
-                cudaSafeCall( cudaGetLastError() );
-            }
-        };
-        template <typename T> struct ResizeDispatcherStream<LinearFilter, T>
+            resize<<<grid, block>>>(filteredSrc, dst, fy, fx);
+        }
+        else
         {
-            static void call(PtrStepSz<T> src, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
-            {
-                const dim3 block(32, 8);
-                const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+            TextureAccessor<T> texSrc = texAccessor(srcWhole, yoff, xoff);
 
-                resize_linear<<<grid, block, 0, stream>>>(src, fx, fy, dst);
-                cudaSafeCall( cudaGetLastError() );
-            }
-        };
+            BrdReplicate<T> brd(src.rows, src.cols);
+            BorderReader<TextureAccessor<T>, BrdReplicate<T> > brdSrc(texSrc, brd);
+            LinearFilter< BorderReader<TextureAccessor<T>, BrdReplicate<T> > > filteredSrc(brdSrc);
 
-        template <template <typename> class Filter, typename T> struct ResizeDispatcherNonStream
-        {
-            static void call(PtrStepSz<T> src, PtrStepSz<T>, int, int, float fx, float fy, PtrStepSz<T> dst)
-            {
-                const dim3 block(32, 8);
-                const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+            resize<<<grid, block>>>(filteredSrc, dst, fy, fx);
+        }
 
-                BrdReplicate<T> brd(src.rows, src.cols);
-                BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
-                Filter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filteredSrc(brdSrc);
+        cudaSafeCall( cudaGetLastError() );
 
-                resize<<<grid, block>>>(filteredSrc, fx, fy, dst);
-                cudaSafeCall( cudaGetLastError() );
+        cudaSafeCall( cudaDeviceSynchronize() );
+    }
 
-                cudaSafeCall( cudaDeviceSynchronize() );
-            }
-        };
-        template <typename T> struct ResizeDispatcherNonStream<PointFilter, T>
-        {
-        static void call(PtrStepSz<T> src, PtrStepSz<T>, int, int, float fx, float fy, PtrStepSz<T> dst)
-            {
-                const dim3 block(32, 8);
-                const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+    // callers for cubic interpolation
 
-                resize_nearest<<<grid, block>>>(src, fx, fy, dst);
-                cudaSafeCall( cudaGetLastError() );
+    template <typename T>
+    void call_resize_cubic_glob(const PtrStepSz<T>& src, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
+    {
+        const dim3 block(32, 8);
+        const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
 
-                cudaSafeCall( cudaDeviceSynchronize() );
-            }
-        };
-        template <typename T> struct ResizeDispatcherNonStream<LinearFilter, T>
+        BrdReplicate<T> brd(src.rows, src.cols);
+        BorderReader< PtrStep<T>, BrdReplicate<T> > brdSrc(src, brd);
+        CubicFilter< BorderReader< PtrStep<T>, BrdReplicate<T> > > filteredSrc(brdSrc);
+
+        resize<<<grid, block, 0, stream>>>(filteredSrc, dst, fy, fx);
+        cudaSafeCall( cudaGetLastError() );
+
+        if (stream == 0)
+            cudaSafeCall( cudaDeviceSynchronize() );
+    }
+
+    template <typename T>
+    void call_resize_cubic_tex(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx)
+    {
+        const dim3 block(32, 8);
+        const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+
+        if (srcWhole.data == src.data)
         {
-        static void call(PtrStepSz<T> src, PtrStepSz<T>, int, int, float fx, float fy, PtrStepSz<T> dst)
-            {
-                const dim3 block(32, 8);
-                const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
+            TextureAccessor<T> texSrc = texAccessor(src, 0, 0);
+            CubicFilter< TextureAccessor<T> > filteredSrc(texSrc);
 
-                resize_linear<<<grid, block>>>(src, fx, fy, dst);
-                cudaSafeCall( cudaGetLastError() );
+            resize<<<grid, block>>>(filteredSrc, dst, fy, fx);
+        }
+        else
+        {
+            TextureAccessor<T> texSrc = texAccessor(srcWhole, yoff, xoff);
 
-                cudaSafeCall( cudaDeviceSynchronize() );
-            }
-        };
+            BrdReplicate<T> brd(src.rows, src.cols);
+            BorderReader<TextureAccessor<T>, BrdReplicate<T> > brdSrc(texSrc, brd);
+            CubicFilter< BorderReader<TextureAccessor<T>, BrdReplicate<T> > > filteredSrc(brdSrc);
 
-        #define OPENCV_GPU_IMPLEMENT_RESIZE_TEX(type) \
-            texture< type , cudaTextureType2D> tex_resize_ ## type (0, cudaFilterModePoint, cudaAddressModeClamp); \
-            struct tex_resize_ ## type ## _reader \
-            { \
-                typedef type elem_type; \
-                typedef int index_type; \
-                int xoff; \
-                int yoff; \
-                __device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const \
-                { \
-                    return tex2D(tex_resize_ ## type, x + xoff, y + yoff); \
-                } \
-            }; \
-            template <template <typename> class Filter> struct ResizeDispatcherNonStream<Filter, type > \
-            { \
-                static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz< type > dst) \
-                { \
-                    const dim3 block(32, 8); \
-                    const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
-                    if (srcWhole.data == src.data) \
-                    { \
-                        bindTexture(&tex_resize_ ## type, src); \
-                        tex_resize_ ## type ## _reader texSrc; \
-                        texSrc.xoff = 0; \
-                        texSrc.yoff = 0; \
-                        Filter<tex_resize_ ## type ## _reader> filteredSrc(texSrc); \
-                        resize<<<grid, block>>>(filteredSrc, fx, fy, dst); \
-                    } \
-                    else \
-                    { \
-                        bindTexture(&tex_resize_ ## type, srcWhole); \
-                        tex_resize_ ## type ## _reader texSrc; \
-                        texSrc.xoff = xoff; \
-                        texSrc.yoff = yoff; \
-                        BrdReplicate< type > brd(src.rows, src.cols); \
-                        BorderReader<tex_resize_ ## type ## _reader, BrdReplicate< type > > brdSrc(texSrc, brd); \
-                        Filter< BorderReader<tex_resize_ ## type ## _reader, BrdReplicate< type > > > filteredSrc(brdSrc); \
-                        resize<<<grid, block>>>(filteredSrc, fx, fy, dst); \
-                    } \
-                    cudaSafeCall( cudaGetLastError() ); \
-                    cudaSafeCall( cudaDeviceSynchronize() ); \
-                } \
-            }; \
-            template <> struct ResizeDispatcherNonStream<PointFilter, type > \
-            { \
-                static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz< type > dst) \
-                { \
-                    const dim3 block(32, 8); \
-                    const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
-                    bindTexture(&tex_resize_ ## type, srcWhole); \
-                    tex_resize_ ## type ## _reader texSrc; \
-                    texSrc.xoff = xoff; \
-                    texSrc.yoff = yoff; \
-                    resize<<<grid, block>>>(texSrc, fx, fy, dst); \
-                    cudaSafeCall( cudaGetLastError() ); \
-                    cudaSafeCall( cudaDeviceSynchronize() ); \
-                } \
-            }; \
-            template <> struct ResizeDispatcherNonStream<LinearFilter, type > \
-            { \
-                static void call(PtrStepSz< type > src, PtrStepSz< type > srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz< type > dst) \
-                { \
-                    const dim3 block(32, 8); \
-                    const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y)); \
-                    if (srcWhole.data == src.data) \
-                    { \
-                        bindTexture(&tex_resize_ ## type, src); \
-                        tex_resize_ ## type ## _reader texSrc; \
-                        texSrc.xoff = 0; \
-                        texSrc.yoff = 0; \
-                        LinearFilter<tex_resize_ ## type ## _reader> filteredSrc(texSrc); \
-                        resize<<<grid, block>>>(filteredSrc, fx, fy, dst); \
-                    } \
-                    else \
-                    { \
-                        bindTexture(&tex_resize_ ## type, srcWhole); \
-                        tex_resize_ ## type ## _reader texSrc; \
-                        texSrc.xoff = xoff; \
-                        texSrc.yoff = yoff; \
-                        BrdReplicate< type > brd(src.rows, src.cols); \
-                        BorderReader<tex_resize_ ## type ## _reader, BrdReplicate< type > > brdSrc(texSrc, brd); \
-                        LinearFilter< BorderReader<tex_resize_ ## type ## _reader, BrdReplicate< type > > > filteredSrc(brdSrc); \
-                        resize<<<grid, block>>>(filteredSrc, fx, fy, dst); \
-                    } \
-                    cudaSafeCall( cudaGetLastError() ); \
-                    cudaSafeCall( cudaDeviceSynchronize() ); \
-                } \
-            };
-
-        OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar)
-        OPENCV_GPU_IMPLEMENT_RESIZE_TEX(uchar4)
-
-        //OPENCV_GPU_IMPLEMENT_RESIZE_TEX(schar)
-        //OPENCV_GPU_IMPLEMENT_RESIZE_TEX(char4)
-
-        OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort)
-        OPENCV_GPU_IMPLEMENT_RESIZE_TEX(ushort4)
-
-        OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short)
-        OPENCV_GPU_IMPLEMENT_RESIZE_TEX(short4)
-
-        //OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int)
-        //OPENCV_GPU_IMPLEMENT_RESIZE_TEX(int4)
-
-        OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float)
-        OPENCV_GPU_IMPLEMENT_RESIZE_TEX(float4)
-
-        #undef OPENCV_GPU_IMPLEMENT_RESIZE_TEX
-
-        template <template <typename> class Filter, typename T> struct ResizeDispatcher
+            resize<<<grid, block>>>(filteredSrc, dst, fy, fx);
+        }
+
+        cudaSafeCall( cudaGetLastError() );
+
+        cudaSafeCall( cudaDeviceSynchronize() );
+    }
+
+    // ResizeNearestDispatcher
+
+    template <typename T> struct ResizeNearestDispatcher
+    {
+        static void call(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
+        {
+            call_resize_nearest_glob(src, dst, fy, fx, stream);
+        }
+    };
+
+    template <typename T> struct SelectImplForNearest
+    {
+        static void call(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
         {
-            static void call(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
+            if (stream)
+                call_resize_nearest_glob(src, dst, fy, fx, stream);
+            else
             {
-                if (stream == 0)
-                    ResizeDispatcherNonStream<Filter, T>::call(src, srcWhole, xoff, yoff, fx, fy, dst);
+                if (fx > 1 || fy > 1)
+                    call_resize_nearest_glob(src, dst, fy, fx, 0);
                 else
-                    ResizeDispatcherStream<Filter, T>::call(src, fx, fy, dst, stream);
+                    call_resize_nearest_tex(src, srcWhole, yoff, xoff, dst, fy, fx);
             }
-        };
+        }
+    };
 
-        template <typename Ptr2D, typename T> __global__ void resize_area(const Ptr2D src, PtrStepSz<T> dst)
-        {
-            const int x = blockDim.x * blockIdx.x + threadIdx.x;
-            const int y = blockDim.y * blockIdx.y + threadIdx.y;
+    template <> struct ResizeNearestDispatcher<uchar> : SelectImplForNearest<uchar> {};
+    template <> struct ResizeNearestDispatcher<uchar4> : SelectImplForNearest<uchar4> {};
 
-            if (x < dst.cols && y < dst.rows)
-            {
-                dst(y, x) = src(y, x);
-            }
+    template <> struct ResizeNearestDispatcher<ushort> : SelectImplForNearest<ushort> {};
+    template <> struct ResizeNearestDispatcher<ushort4> : SelectImplForNearest<ushort4> {};
+
+    template <> struct ResizeNearestDispatcher<short> : SelectImplForNearest<short> {};
+    template <> struct ResizeNearestDispatcher<short4> : SelectImplForNearest<short4> {};
+
+    template <> struct ResizeNearestDispatcher<float> : SelectImplForNearest<float> {};
+    template <> struct ResizeNearestDispatcher<float4> : SelectImplForNearest<float4> {};
+
+    // ResizeLinearDispatcher
+
+    template <typename T> struct ResizeLinearDispatcher
+    {
+        static void call(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
+        {
+            call_resize_linear_glob(src, dst, fy, fx, stream);
         }
+    };
 
-        template <typename T> struct ResizeAreaDispatcher
+    template <typename T> struct SelectImplForLinear
+    {
+        static void call(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
         {
-            static void call(PtrStepSz<T> src, PtrStepSz<T>, int, int, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream)
+            if (stream)
+                call_resize_linear_glob(src, dst, fy, fx, stream);
+            else
             {
-                const int iscale_x = (int) round(fx);
-                const int iscale_y = (int) round(fy);
-
-                const dim3 block(32, 8);
-                const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
-
-                if (std::abs(fx - iscale_x) < FLT_MIN && std::abs(fy - iscale_y) < FLT_MIN)
-                {
-                    BrdConstant<T> brd(src.rows, src.cols);
-                    BorderReader< PtrStep<T>, BrdConstant<T> > brdSrc(src, brd);
-                    IntegerAreaFilter< BorderReader< PtrStep<T>, BrdConstant<T> > > filteredSrc(brdSrc, fx, fy);
-                    resize_area<<<grid, block, 0, stream>>>(filteredSrc, dst);
-                }
+                if (fx > 1 || fy > 1)
+                    call_resize_linear_glob(src, dst, fy, fx, 0);
                 else
-                {
-                    BrdConstant<T> brd(src.rows, src.cols);
-                    BorderReader< PtrStep<T>, BrdConstant<T> > brdSrc(src, brd);
-                    AreaFilter< BorderReader< PtrStep<T>, BrdConstant<T> > > filteredSrc(brdSrc, fx, fy);
-                    resize_area<<<grid, block, 0, stream>>>(filteredSrc, dst);
-                }
-
-                cudaSafeCall( cudaGetLastError() );
-                if (stream == 0)
-                    cudaSafeCall( cudaDeviceSynchronize() );
+                    call_resize_linear_tex(src, srcWhole, yoff, xoff, dst, fy, fx);
             }
-        };
+        }
+    };
+
+    template <> struct ResizeLinearDispatcher<uchar> : SelectImplForLinear<uchar> {};
+    template <> struct ResizeLinearDispatcher<uchar4> : SelectImplForLinear<uchar4> {};
+
+    template <> struct ResizeLinearDispatcher<ushort> : SelectImplForLinear<ushort> {};
+    template <> struct ResizeLinearDispatcher<ushort4> : SelectImplForLinear<ushort4> {};
 
-        template <typename T> void resize_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy,
-            PtrStepSzb dst, int interpolation, cudaStream_t stream)
+    template <> struct ResizeLinearDispatcher<short> : SelectImplForLinear<short> {};
+    template <> struct ResizeLinearDispatcher<short4> : SelectImplForLinear<short4> {};
+
+    template <> struct ResizeLinearDispatcher<float> : SelectImplForLinear<float> {};
+    template <> struct ResizeLinearDispatcher<float4> : SelectImplForLinear<float4> {};
+
+    // ResizeCubicDispatcher
+
+    template <typename T> struct ResizeCubicDispatcher
+    {
+        static void call(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
+        {
+            call_resize_cubic_glob(src, dst, fy, fx, stream);
+        }
+    };
+
+    template <typename T> struct SelectImplForCubic
+    {
+        static void call(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
+        {
+            if (stream)
+                call_resize_cubic_glob(src, dst, fy, fx, stream);
+            else
+                call_resize_cubic_tex(src, srcWhole, yoff, xoff, dst, fy, fx);
+        }
+    };
+
+    template <> struct ResizeCubicDispatcher<uchar> : SelectImplForCubic<uchar> {};
+    template <> struct ResizeCubicDispatcher<uchar4> : SelectImplForCubic<uchar4> {};
+
+    template <> struct ResizeCubicDispatcher<ushort> : SelectImplForCubic<ushort> {};
+    template <> struct ResizeCubicDispatcher<ushort4> : SelectImplForCubic<ushort4> {};
+
+    template <> struct ResizeCubicDispatcher<short> : SelectImplForCubic<short> {};
+    template <> struct ResizeCubicDispatcher<short4> : SelectImplForCubic<short4> {};
+
+    template <> struct ResizeCubicDispatcher<float> : SelectImplForCubic<float> {};
+    template <> struct ResizeCubicDispatcher<float4> : SelectImplForCubic<float4> {};
+
+    // ResizeAreaDispatcher
+
+    template <typename T> struct ResizeAreaDispatcher
+    {
+        static void call(const PtrStepSz<T>& src, const PtrStepSz<T>&, int, int, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream)
         {
-            typedef void (*caller_t)(PtrStepSz<T> src, PtrStepSz<T> srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSz<T> dst, cudaStream_t stream);
+            const int iscale_x = (int) round(fx);
+            const int iscale_y = (int) round(fy);
+
+            const dim3 block(32, 8);
+            const dim3 grid(divUp(dst.cols, block.x), divUp(dst.rows, block.y));
 
-            static const caller_t callers[4] =
+            if (std::abs(fx - iscale_x) < FLT_MIN && std::abs(fy - iscale_y) < FLT_MIN)
             {
-                ResizeDispatcher<PointFilter, T>::call,
-                ResizeDispatcher<LinearFilter, T>::call,
-                ResizeDispatcher<CubicFilter, T>::call,
-                ResizeAreaDispatcher<T>::call
-            };
-
-            // change to linear if area interpolation upscaling
-            if (interpolation == 3 && (fx <= 1.f || fy <= 1.f))
-                interpolation = 1;
-
-            callers[interpolation](static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(srcWhole), xoff, yoff, fx, fy,
-                static_cast< PtrStepSz<T> >(dst), stream);
+                BrdConstant<T> brd(src.rows, src.cols);
+                BorderReader< PtrStep<T>, BrdConstant<T> > brdSrc(src, brd);
+                IntegerAreaFilter< BorderReader< PtrStep<T>, BrdConstant<T> > > filteredSrc(brdSrc, fx, fy);
+
+                resize_area<<<grid, block, 0, stream>>>(filteredSrc, dst);
+            }
+            else
+            {
+                BrdConstant<T> brd(src.rows, src.cols);
+                BorderReader< PtrStep<T>, BrdConstant<T> > brdSrc(src, brd);
+                AreaFilter< BorderReader< PtrStep<T>, BrdConstant<T> > > filteredSrc(brdSrc, fx, fy);
+
+                resize_area<<<grid, block, 0, stream>>>(filteredSrc, dst);
+            }
+
+            cudaSafeCall( cudaGetLastError() );
+
+            if (stream == 0)
+                cudaSafeCall( cudaDeviceSynchronize() );
         }
+    };
+
+    // resize
+
+    template <typename T> void resize(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream)
+    {
+        typedef void (*func_t)(const PtrStepSz<T>& src, const PtrStepSz<T>& srcWhole, int yoff, int xoff, const PtrStepSz<T>& dst, float fy, float fx, cudaStream_t stream);
+        static const func_t funcs[4] =
+        {
+            ResizeNearestDispatcher<T>::call,
+            ResizeLinearDispatcher<T>::call,
+            ResizeCubicDispatcher<T>::call,
+            ResizeAreaDispatcher<T>::call
+        };
+
+        // change to linear if area interpolation upscaling
+        if (interpolation == 3 && (fx <= 1.f || fy <= 1.f))
+            interpolation = 1;
+
+        funcs[interpolation](static_cast< PtrStepSz<T> >(src), static_cast< PtrStepSz<T> >(srcWhole), yoff, xoff, static_cast< PtrStepSz<T> >(dst), fy, fx, stream);
+    }
+
+    template void resize<uchar >(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+    template void resize<uchar3>(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+    template void resize<uchar4>(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+
+    template void resize<ushort >(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+    template void resize<ushort3>(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+    template void resize<ushort4>(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+
+    template void resize<short >(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+    template void resize<short3>(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+    template void resize<short4>(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
 
-        template void resize_gpu<uchar >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<uchar2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        template void resize_gpu<uchar3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        template void resize_gpu<uchar4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-
-        //template void resize_gpu<schar>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<char2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<char3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<char4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-
-        template void resize_gpu<ushort >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<ushort2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        template void resize_gpu<ushort3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        template void resize_gpu<ushort4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-
-        template void resize_gpu<short >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<short2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        template void resize_gpu<short3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        template void resize_gpu<short4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-
-        //template void resize_gpu<int >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<int2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<int3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<int4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-
-        template void resize_gpu<float >(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        //template void resize_gpu<float2>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        template void resize_gpu<float3>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-        template void resize_gpu<float4>(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
-    } // namespace imgproc
-}}} // namespace cv { namespace gpu { namespace device
+    template void resize<float >(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+    template void resize<float3>(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+    template void resize<float4>(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
+}}}
 
 #endif /* CUDA_DISABLER */
diff --git a/modules/gpu/src/resize.cpp b/modules/gpu/src/resize.cpp
index 0e8c865..e1b5026 100644
--- a/modules/gpu/src/resize.cpp
+++ b/modules/gpu/src/resize.cpp
@@ -50,27 +50,21 @@ void cv::gpu::resize(const GpuMat&, GpuMat&, Size, double, double, int, Stream&)
 
 namespace cv { namespace gpu { namespace device
 {
-    namespace imgproc
-    {
-        template <typename T>
-        void resize_gpu(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy,
-                        PtrStepSzb dst, int interpolation, cudaStream_t stream);
-    }
+    template <typename T>
+    void resize(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
 }}}
 
 void cv::gpu::resize(const GpuMat& src, GpuMat& dst, Size dsize, double fx, double fy, int interpolation, Stream& stream)
 {
-    using namespace ::cv::gpu::device::imgproc;
-
-    typedef void (*func_t)(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float fx, float fy, PtrStepSzb dst, int interpolation, cudaStream_t stream);
+    typedef void (*func_t)(const PtrStepSzb& src, const PtrStepSzb& srcWhole, int yoff, int xoff, const PtrStepSzb& dst, float fy, float fx, int interpolation, cudaStream_t stream);
     static const func_t funcs[6][4] =
     {
-        {resize_gpu<uchar>      , 0 /*resize_gpu<uchar2>*/ , resize_gpu<uchar3>     , resize_gpu<uchar4>     },
-        {0 /*resize_gpu<schar>*/, 0 /*resize_gpu<char2>*/  , 0 /*resize_gpu<char3>*/, 0 /*resize_gpu<char4>*/},
-        {resize_gpu<ushort>     , 0 /*resize_gpu<ushort2>*/, resize_gpu<ushort3>    , resize_gpu<ushort4>    },
-        {resize_gpu<short>      , 0 /*resize_gpu<short2>*/ , resize_gpu<short3>     , resize_gpu<short4>     },
-        {0 /*resize_gpu<int>*/  , 0 /*resize_gpu<int2>*/   , 0 /*resize_gpu<int3>*/ , 0 /*resize_gpu<int4>*/ },
-        {resize_gpu<float>      , 0 /*resize_gpu<float2>*/ , resize_gpu<float3>     , resize_gpu<float4>     }
+        {device::resize<uchar>      , 0 /*device::resize<uchar2>*/ , device::resize<uchar3>     , device::resize<uchar4>     },
+        {0 /*device::resize<schar>*/, 0 /*device::resize<char2>*/  , 0 /*device::resize<char3>*/, 0 /*device::resize<char4>*/},
+        {device::resize<ushort>     , 0 /*device::resize<ushort2>*/, device::resize<ushort3>    , device::resize<ushort4>    },
+        {device::resize<short>      , 0 /*device::resize<short2>*/ , device::resize<short3>     , device::resize<short4>     },
+        {0 /*device::resize<int>*/  , 0 /*device::resize<int2>*/   , 0 /*device::resize<int3>*/ , 0 /*device::resize<int4>*/ },
+        {device::resize<float>      , 0 /*device::resize<float2>*/ , device::resize<float3>     , device::resize<float4>     }
     };
 
     CV_Assert( src.depth() <= CV_32F && src.channels() <= 4 );
@@ -108,7 +102,7 @@ void cv::gpu::resize(const GpuMat& src, GpuMat& dst, Size dsize, double fx, doub
     src.locateROI(wholeSize, ofs);
     PtrStepSzb wholeSrc(wholeSize.height, wholeSize.width, src.datastart, src.step);
 
-    func(src, wholeSrc, ofs.x, ofs.y, static_cast<float>(1.0 / fx), static_cast<float>(1.0 / fy), dst, interpolation, StreamAccessor::getStream(stream));
+    func(src, wholeSrc, ofs.y, ofs.x, dst, static_cast<float>(1.0 / fy), static_cast<float>(1.0 / fx), interpolation, StreamAccessor::getStream(stream));
 }
 
 #endif // HAVE_CUDA
-- 
2.7.4