Initial commit.

Final release of the project Anonymizer (2015).
Project settings for the Qt Creator (ver. 3.6).

3rdparty/include/opencv2/core/cuda/block.hpp

@@ -0,0 +1,211 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_DEVICE_BLOCK_HPP__
+#define __OPENCV_CUDA_DEVICE_BLOCK_HPP__
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    struct Block
+    {
+        static __device__ __forceinline__ unsigned int id()
+        {
+            return blockIdx.x;
+        }
+
+        static __device__ __forceinline__ unsigned int stride()
+        {
+            return blockDim.x * blockDim.y * blockDim.z;
+        }
+
+        static __device__ __forceinline__ void sync()
+        {
+            __syncthreads();
+        }
+
+        static __device__ __forceinline__ int flattenedThreadId()
+        {
+            return threadIdx.z * blockDim.x * blockDim.y + threadIdx.y * blockDim.x + threadIdx.x;
+        }
+
+        template<typename It, typename T>
+        static __device__ __forceinline__ void fill(It beg, It end, const T& value)
+        {
+            int STRIDE = stride();
+            It t = beg + flattenedThreadId();
+
+            for(; t < end; t += STRIDE)
+                *t = value;
+        }
+
+        template<typename OutIt, typename T>
+        static __device__ __forceinline__ void yota(OutIt beg, OutIt end, T value)
+        {
+            int STRIDE = stride();
+            int tid = flattenedThreadId();
+            value += tid;
+
+            for(OutIt t = beg + tid; t < end; t += STRIDE, value += STRIDE)
+                *t = value;
+        }
+
+        template<typename InIt, typename OutIt>
+        static __device__ __forceinline__ void copy(InIt beg, InIt end, OutIt out)
+        {
+            int STRIDE = stride();
+            InIt  t = beg + flattenedThreadId();
+            OutIt o = out + (t - beg);
+
+            for(; t < end; t += STRIDE, o += STRIDE)
+                *o = *t;
+        }
+
+        template<typename InIt, typename OutIt, class UnOp>
+        static __device__ __forceinline__ void transfrom(InIt beg, InIt end, OutIt out, UnOp op)
+        {
+            int STRIDE = stride();
+            InIt  t = beg + flattenedThreadId();
+            OutIt o = out + (t - beg);
+
+            for(; t < end; t += STRIDE, o += STRIDE)
+                *o = op(*t);
+        }
+
+        template<typename InIt1, typename InIt2, typename OutIt, class BinOp>
+        static __device__ __forceinline__ void transfrom(InIt1 beg1, InIt1 end1, InIt2 beg2, OutIt out, BinOp op)
+        {
+            int STRIDE = stride();
+            InIt1 t1 = beg1 + flattenedThreadId();
+            InIt2 t2 = beg2 + flattenedThreadId();
+            OutIt o  = out + (t1 - beg1);
+
+            for(; t1 < end1; t1 += STRIDE, t2 += STRIDE, o += STRIDE)
+                *o = op(*t1, *t2);
+        }
+
+        template<int CTA_SIZE, typename T, class BinOp>
+        static __device__ __forceinline__ void reduce(volatile T* buffer, BinOp op)
+        {
+            int tid = flattenedThreadId();
+            T val =  buffer[tid];
+
+            if (CTA_SIZE >= 1024) { if (tid < 512) buffer[tid] = val = op(val, buffer[tid + 512]); __syncthreads(); }
+            if (CTA_SIZE >=  512) { if (tid < 256) buffer[tid] = val = op(val, buffer[tid + 256]); __syncthreads(); }
+            if (CTA_SIZE >=  256) { if (tid < 128) buffer[tid] = val = op(val, buffer[tid + 128]); __syncthreads(); }
+            if (CTA_SIZE >=  128) { if (tid <  64) buffer[tid] = val = op(val, buffer[tid +  64]); __syncthreads(); }
+
+            if (tid < 32)
+            {
+                if (CTA_SIZE >=   64) { buffer[tid] = val = op(val, buffer[tid +  32]); }
+                if (CTA_SIZE >=   32) { buffer[tid] = val = op(val, buffer[tid +  16]); }
+                if (CTA_SIZE >=   16) { buffer[tid] = val = op(val, buffer[tid +   8]); }
+                if (CTA_SIZE >=    8) { buffer[tid] = val = op(val, buffer[tid +   4]); }
+                if (CTA_SIZE >=    4) { buffer[tid] = val = op(val, buffer[tid +   2]); }
+                if (CTA_SIZE >=    2) { buffer[tid] = val = op(val, buffer[tid +   1]); }
+            }
+        }
+
+        template<int CTA_SIZE, typename T, class BinOp>
+        static __device__ __forceinline__ T reduce(volatile T* buffer, T init, BinOp op)
+        {
+            int tid = flattenedThreadId();
+            T val =  buffer[tid] = init;
+            __syncthreads();
+
+            if (CTA_SIZE >= 1024) { if (tid < 512) buffer[tid] = val = op(val, buffer[tid + 512]); __syncthreads(); }
+            if (CTA_SIZE >=  512) { if (tid < 256) buffer[tid] = val = op(val, buffer[tid + 256]); __syncthreads(); }
+            if (CTA_SIZE >=  256) { if (tid < 128) buffer[tid] = val = op(val, buffer[tid + 128]); __syncthreads(); }
+            if (CTA_SIZE >=  128) { if (tid <  64) buffer[tid] = val = op(val, buffer[tid +  64]); __syncthreads(); }
+
+            if (tid < 32)
+            {
+                if (CTA_SIZE >=   64) { buffer[tid] = val = op(val, buffer[tid +  32]); }
+                if (CTA_SIZE >=   32) { buffer[tid] = val = op(val, buffer[tid +  16]); }
+                if (CTA_SIZE >=   16) { buffer[tid] = val = op(val, buffer[tid +   8]); }
+                if (CTA_SIZE >=    8) { buffer[tid] = val = op(val, buffer[tid +   4]); }
+                if (CTA_SIZE >=    4) { buffer[tid] = val = op(val, buffer[tid +   2]); }
+                if (CTA_SIZE >=    2) { buffer[tid] = val = op(val, buffer[tid +   1]); }
+            }
+            __syncthreads();
+            return buffer[0];
+        }
+
+        template <typename T, class BinOp>
+        static __device__ __forceinline__ void reduce_n(T* data, unsigned int n, BinOp op)
+        {
+            int ftid = flattenedThreadId();
+            int sft = stride();
+
+            if (sft < n)
+            {
+                for (unsigned int i = sft + ftid; i < n; i += sft)
+                    data[ftid] = op(data[ftid], data[i]);
+
+                __syncthreads();
+
+                n = sft;
+            }
+
+            while (n > 1)
+            {
+                unsigned int half = n/2;
+
+                if (ftid < half)
+                    data[ftid] = op(data[ftid], data[n - ftid - 1]);
+
+                __syncthreads();
+
+                n = n - half;
+            }
+        }
+    };
+}}}
+
+//! @endcond
+
+#endif /* __OPENCV_CUDA_DEVICE_BLOCK_HPP__ */

3rdparty/include/opencv2/core/cuda/border_interpolate.hpp

@@ -0,0 +1,722 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_BORDER_INTERPOLATE_HPP__
+#define __OPENCV_CUDA_BORDER_INTERPOLATE_HPP__
+
+#include "saturate_cast.hpp"
+#include "vec_traits.hpp"
+#include "vec_math.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    //////////////////////////////////////////////////////////////
+    // BrdConstant
+
+    template <typename D> struct BrdRowConstant
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdRowConstant(int width_, const D& val_ = VecTraits<D>::all(0)) : width(width_), val(val_) {}
+
+        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
+        {
+            return x >= 0 ? saturate_cast<D>(data[x]) : val;
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
+        {
+            return x < width ? saturate_cast<D>(data[x]) : val;
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
+        {
+            return (x >= 0 && x < width) ? saturate_cast<D>(data[x]) : val;
+        }
+
+        int width;
+        D val;
+    };
+
+    template <typename D> struct BrdColConstant
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdColConstant(int height_, const D& val_ = VecTraits<D>::all(0)) : height(height_), val(val_) {}
+
+        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
+        {
+            return y >= 0 ? saturate_cast<D>(*(const T*)((const char*)data + y * step)) : val;
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
+        {
+            return y < height ? saturate_cast<D>(*(const T*)((const char*)data + y * step)) : val;
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
+        {
+            return (y >= 0 && y < height) ? saturate_cast<D>(*(const T*)((const char*)data + y * step)) : val;
+        }
+
+        int height;
+        D val;
+    };
+
+    template <typename D> struct BrdConstant
+    {
+        typedef D result_type;
+
+        __host__ __device__ __forceinline__ BrdConstant(int height_, int width_, const D& val_ = VecTraits<D>::all(0)) : height(height_), width(width_), val(val_)
+        {
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
+        {
+            return (x >= 0 && x < width && y >= 0 && y < height) ? saturate_cast<D>(((const T*)((const uchar*)data + y * step))[x]) : val;
+        }
+
+        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
+        {
+            return (x >= 0 && x < width && y >= 0 && y < height) ? saturate_cast<D>(src(y, x)) : val;
+        }
+
+        int height;
+        int width;
+        D val;
+    };
+
+    //////////////////////////////////////////////////////////////
+    // BrdReplicate
+
+    template <typename D> struct BrdRowReplicate
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdRowReplicate(int width) : last_col(width - 1) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdRowReplicate(int width, U) : last_col(width - 1) {}
+
+        __device__ __forceinline__ int idx_col_low(int x) const
+        {
+            return ::max(x, 0);
+        }
+
+        __device__ __forceinline__ int idx_col_high(int x) const
+        {
+            return ::min(x, last_col);
+        }
+
+        __device__ __forceinline__ int idx_col(int x) const
+        {
+            return idx_col_low(idx_col_high(x));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col_low(x)]);
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col_high(x)]);
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col(x)]);
+        }
+
+        int last_col;
+    };
+
+    template <typename D> struct BrdColReplicate
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdColReplicate(int height) : last_row(height - 1) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdColReplicate(int height, U) : last_row(height - 1) {}
+
+        __device__ __forceinline__ int idx_row_low(int y) const
+        {
+            return ::max(y, 0);
+        }
+
+        __device__ __forceinline__ int idx_row_high(int y) const
+        {
+            return ::min(y, last_row);
+        }
+
+        __device__ __forceinline__ int idx_row(int y) const
+        {
+            return idx_row_low(idx_row_high(y));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const T*)((const char*)data + idx_row_low(y) * step));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const T*)((const char*)data + idx_row_high(y) * step));
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const T*)((const char*)data + idx_row(y) * step));
+        }
+
+        int last_row;
+    };
+
+    template <typename D> struct BrdReplicate
+    {
+        typedef D result_type;
+
+        __host__ __device__ __forceinline__ BrdReplicate(int height, int width) : last_row(height - 1), last_col(width - 1) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdReplicate(int height, int width, U) : last_row(height - 1), last_col(width - 1) {}
+
+        __device__ __forceinline__ int idx_row_low(int y) const
+        {
+            return ::max(y, 0);
+        }
+
+        __device__ __forceinline__ int idx_row_high(int y) const
+        {
+            return ::min(y, last_row);
+        }
+
+        __device__ __forceinline__ int idx_row(int y) const
+        {
+            return idx_row_low(idx_row_high(y));
+        }
+
+        __device__ __forceinline__ int idx_col_low(int x) const
+        {
+            return ::max(x, 0);
+        }
+
+        __device__ __forceinline__ int idx_col_high(int x) const
+        {
+            return ::min(x, last_col);
+        }
+
+        __device__ __forceinline__ int idx_col(int x) const
+        {
+            return idx_col_low(idx_col_high(x));
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(((const T*)((const char*)data + idx_row(y) * step))[idx_col(x)]);
+        }
+
+        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
+        {
+            return saturate_cast<D>(src(idx_row(y), idx_col(x)));
+        }
+
+        int last_row;
+        int last_col;
+    };
+
+    //////////////////////////////////////////////////////////////
+    // BrdReflect101
+
+    template <typename D> struct BrdRowReflect101
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdRowReflect101(int width) : last_col(width - 1) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdRowReflect101(int width, U) : last_col(width - 1) {}
+
+        __device__ __forceinline__ int idx_col_low(int x) const
+        {
+            return ::abs(x) % (last_col + 1);
+        }
+
+        __device__ __forceinline__ int idx_col_high(int x) const
+        {
+            return ::abs(last_col - ::abs(last_col - x)) % (last_col + 1);
+        }
+
+        __device__ __forceinline__ int idx_col(int x) const
+        {
+            return idx_col_low(idx_col_high(x));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col_low(x)]);
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col_high(x)]);
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col(x)]);
+        }
+
+        int last_col;
+    };
+
+    template <typename D> struct BrdColReflect101
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdColReflect101(int height) : last_row(height - 1) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdColReflect101(int height, U) : last_row(height - 1) {}
+
+        __device__ __forceinline__ int idx_row_low(int y) const
+        {
+            return ::abs(y) % (last_row + 1);
+        }
+
+        __device__ __forceinline__ int idx_row_high(int y) const
+        {
+            return ::abs(last_row - ::abs(last_row - y)) % (last_row + 1);
+        }
+
+        __device__ __forceinline__ int idx_row(int y) const
+        {
+            return idx_row_low(idx_row_high(y));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_low(y) * step));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_high(y) * step));
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const D*)((const char*)data + idx_row(y) * step));
+        }
+
+        int last_row;
+    };
+
+    template <typename D> struct BrdReflect101
+    {
+        typedef D result_type;
+
+        __host__ __device__ __forceinline__ BrdReflect101(int height, int width) : last_row(height - 1), last_col(width - 1) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdReflect101(int height, int width, U) : last_row(height - 1), last_col(width - 1) {}
+
+        __device__ __forceinline__ int idx_row_low(int y) const
+        {
+            return ::abs(y) % (last_row + 1);
+        }
+
+        __device__ __forceinline__ int idx_row_high(int y) const
+        {
+            return ::abs(last_row - ::abs(last_row - y)) % (last_row + 1);
+        }
+
+        __device__ __forceinline__ int idx_row(int y) const
+        {
+            return idx_row_low(idx_row_high(y));
+        }
+
+        __device__ __forceinline__ int idx_col_low(int x) const
+        {
+            return ::abs(x) % (last_col + 1);
+        }
+
+        __device__ __forceinline__ int idx_col_high(int x) const
+        {
+            return ::abs(last_col - ::abs(last_col - x)) % (last_col + 1);
+        }
+
+        __device__ __forceinline__ int idx_col(int x) const
+        {
+            return idx_col_low(idx_col_high(x));
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(((const T*)((const char*)data + idx_row(y) * step))[idx_col(x)]);
+        }
+
+        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
+        {
+            return saturate_cast<D>(src(idx_row(y), idx_col(x)));
+        }
+
+        int last_row;
+        int last_col;
+    };
+
+    //////////////////////////////////////////////////////////////
+    // BrdReflect
+
+    template <typename D> struct BrdRowReflect
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdRowReflect(int width) : last_col(width - 1) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdRowReflect(int width, U) : last_col(width - 1) {}
+
+        __device__ __forceinline__ int idx_col_low(int x) const
+        {
+            return (::abs(x) - (x < 0)) % (last_col + 1);
+        }
+
+        __device__ __forceinline__ int idx_col_high(int x) const
+        {
+            return ::abs(last_col - ::abs(last_col - x) + (x > last_col)) % (last_col + 1);
+        }
+
+        __device__ __forceinline__ int idx_col(int x) const
+        {
+            return idx_col_high(::abs(x) - (x < 0));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col_low(x)]);
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col_high(x)]);
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col(x)]);
+        }
+
+        int last_col;
+    };
+
+    template <typename D> struct BrdColReflect
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdColReflect(int height) : last_row(height - 1) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdColReflect(int height, U) : last_row(height - 1) {}
+
+        __device__ __forceinline__ int idx_row_low(int y) const
+        {
+            return (::abs(y) - (y < 0)) % (last_row + 1);
+        }
+
+        __device__ __forceinline__ int idx_row_high(int y) const
+        {
+            return ::abs(last_row - ::abs(last_row - y) + (y > last_row)) % (last_row + 1);
+        }
+
+        __device__ __forceinline__ int idx_row(int y) const
+        {
+            return idx_row_high(::abs(y) - (y < 0));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_low(y) * step));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_high(y) * step));
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const D*)((const char*)data + idx_row(y) * step));
+        }
+
+        int last_row;
+    };
+
+    template <typename D> struct BrdReflect
+    {
+        typedef D result_type;
+
+        __host__ __device__ __forceinline__ BrdReflect(int height, int width) : last_row(height - 1), last_col(width - 1) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdReflect(int height, int width, U) : last_row(height - 1), last_col(width - 1) {}
+
+        __device__ __forceinline__ int idx_row_low(int y) const
+        {
+            return (::abs(y) - (y < 0)) % (last_row + 1);
+        }
+
+        __device__ __forceinline__ int idx_row_high(int y) const
+        {
+            return /*::abs*/(last_row - ::abs(last_row - y) + (y > last_row)) /*% (last_row + 1)*/;
+        }
+
+        __device__ __forceinline__ int idx_row(int y) const
+        {
+            return idx_row_low(idx_row_high(y));
+        }
+
+        __device__ __forceinline__ int idx_col_low(int x) const
+        {
+            return (::abs(x) - (x < 0)) % (last_col + 1);
+        }
+
+        __device__ __forceinline__ int idx_col_high(int x) const
+        {
+            return (last_col - ::abs(last_col - x) + (x > last_col));
+        }
+
+        __device__ __forceinline__ int idx_col(int x) const
+        {
+            return idx_col_low(idx_col_high(x));
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(((const T*)((const char*)data + idx_row(y) * step))[idx_col(x)]);
+        }
+
+        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
+        {
+            return saturate_cast<D>(src(idx_row(y), idx_col(x)));
+        }
+
+        int last_row;
+        int last_col;
+    };
+
+    //////////////////////////////////////////////////////////////
+    // BrdWrap
+
+    template <typename D> struct BrdRowWrap
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdRowWrap(int width_) : width(width_) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdRowWrap(int width_, U) : width(width_) {}
+
+        __device__ __forceinline__ int idx_col_low(int x) const
+        {
+            return (x >= 0) * x + (x < 0) * (x - ((x - width + 1) / width) * width);
+        }
+
+        __device__ __forceinline__ int idx_col_high(int x) const
+        {
+            return (x < width) * x + (x >= width) * (x % width);
+        }
+
+        __device__ __forceinline__ int idx_col(int x) const
+        {
+            return idx_col_high(idx_col_low(x));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col_low(x)]);
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col_high(x)]);
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
+        {
+            return saturate_cast<D>(data[idx_col(x)]);
+        }
+
+        int width;
+    };
+
+    template <typename D> struct BrdColWrap
+    {
+        typedef D result_type;
+
+        explicit __host__ __device__ __forceinline__ BrdColWrap(int height_) : height(height_) {}
+        template <typename U> __host__ __device__ __forceinline__ BrdColWrap(int height_, U) : height(height_) {}
+
+        __device__ __forceinline__ int idx_row_low(int y) const
+        {
+            return (y >= 0) * y + (y < 0) * (y - ((y - height + 1) / height) * height);
+        }
+
+        __device__ __forceinline__ int idx_row_high(int y) const
+        {
+            return (y < height) * y + (y >= height) * (y % height);
+        }
+
+        __device__ __forceinline__ int idx_row(int y) const
+        {
+            return idx_row_high(idx_row_low(y));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_low(y) * step));
+        }
+
+        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_high(y) * step));
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(*(const D*)((const char*)data + idx_row(y) * step));
+        }
+
+        int height;
+    };
+
+    template <typename D> struct BrdWrap
+    {
+        typedef D result_type;
+
+        __host__ __device__ __forceinline__ BrdWrap(int height_, int width_) :
+            height(height_), width(width_)
+        {
+        }
+        template <typename U>
+        __host__ __device__ __forceinline__ BrdWrap(int height_, int width_, U) :
+            height(height_), width(width_)
+        {
+        }
+
+        __device__ __forceinline__ int idx_row_low(int y) const
+        {
+            return (y >= 0) * y + (y < 0) * (y - ((y - height + 1) / height) * height);
+        }
+
+        __device__ __forceinline__ int idx_row_high(int y) const
+        {
+            return (y < height) * y + (y >= height) * (y % height);
+        }
+
+        __device__ __forceinline__ int idx_row(int y) const
+        {
+            return idx_row_high(idx_row_low(y));
+        }
+
+        __device__ __forceinline__ int idx_col_low(int x) const
+        {
+            return (x >= 0) * x + (x < 0) * (x - ((x - width + 1) / width) * width);
+        }
+
+        __device__ __forceinline__ int idx_col_high(int x) const
+        {
+            return (x < width) * x + (x >= width) * (x % width);
+        }
+
+        __device__ __forceinline__ int idx_col(int x) const
+        {
+            return idx_col_high(idx_col_low(x));
+        }
+
+        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
+        {
+            return saturate_cast<D>(((const T*)((const char*)data + idx_row(y) * step))[idx_col(x)]);
+        }
+
+        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
+        {
+            return saturate_cast<D>(src(idx_row(y), idx_col(x)));
+        }
+
+        int height;
+        int width;
+    };
+
+    //////////////////////////////////////////////////////////////
+    // BorderReader
+
+    template <typename Ptr2D, typename B> struct BorderReader
+    {
+        typedef typename B::result_type elem_type;
+        typedef typename Ptr2D::index_type index_type;
+
+        __host__ __device__ __forceinline__ BorderReader(const Ptr2D& ptr_, const B& b_) : ptr(ptr_), b(b_) {}
+
+        __device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const
+        {
+            return b.at(y, x, ptr);
+        }
+
+        Ptr2D ptr;
+        B b;
+    };
+
+    // under win32 there is some bug with templated types that passed as kernel parameters
+    // with this specialization all works fine
+    template <typename Ptr2D, typename D> struct BorderReader< Ptr2D, BrdConstant<D> >
+    {
+        typedef typename BrdConstant<D>::result_type elem_type;
+        typedef typename Ptr2D::index_type index_type;
+
+        __host__ __device__ __forceinline__ BorderReader(const Ptr2D& src_, const BrdConstant<D>& b) :
+            src(src_), height(b.height), width(b.width), val(b.val)
+        {
+        }
+
+        __device__ __forceinline__ D operator ()(index_type y, index_type x) const
+        {
+            return (x >= 0 && x < width && y >= 0 && y < height) ? saturate_cast<D>(src(y, x)) : val;
+        }
+
+        Ptr2D src;
+        int height;
+        int width;
+        D val;
+    };
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_BORDER_INTERPOLATE_HPP__

3rdparty/include/opencv2/core/cuda/color.hpp

@@ -0,0 +1,309 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_COLOR_HPP__
+#define __OPENCV_CUDA_COLOR_HPP__
+
+#include "detail/color_detail.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    // All OPENCV_CUDA_IMPLEMENT_*_TRAITS(ColorSpace1_to_ColorSpace2, ...) macros implements
+    // template <typename T> class ColorSpace1_to_ColorSpace2_traits
+    // {
+    //     typedef ... functor_type;
+    //     static __host__ __device__ functor_type create_functor();
+    // };
+
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgb, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_bgra, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgba, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_bgr, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgb, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgba, 4, 4, 2)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr555, 3, 0, 5)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr565, 3, 0, 6)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr555, 3, 2, 5)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr565, 3, 2, 6)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr555, 4, 0, 5)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr565, 4, 0, 6)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr555, 4, 2, 5)
+    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr565, 4, 2, 6)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgb, 3, 2, 5)
+    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgb, 3, 2, 6)
+    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgr, 3, 0, 5)
+    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgr, 3, 0, 6)
+    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgba, 4, 2, 5)
+    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgba, 4, 2, 6)
+    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgra, 4, 0, 5)
+    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgra, 4, 0, 6)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgr, 3)
+    OPENCV_CUDA_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgra, 4)
+
+    #undef OPENCV_CUDA_IMPLEMENT_GRAY2RGB_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr555, 5)
+    OPENCV_CUDA_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr565, 6)
+
+    #undef OPENCV_CUDA_IMPLEMENT_GRAY2RGB5x5_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr555_to_gray, 5)
+    OPENCV_CUDA_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr565_to_gray, 6)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB5x52GRAY_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS(rgb_to_gray, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS(bgr_to_gray, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS(rgba_to_gray, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS(bgra_to_gray, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv4, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv4, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv4, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv4, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgb, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgba, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgb, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgba, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgr, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgra, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgr, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgra, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb4, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb4, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb4, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb4, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgb, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgba, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgb, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgba, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgr, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgra, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgr, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgra, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz4, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz4, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz4, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz4, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgb, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgb, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgba, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgba, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgr, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgr, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgra, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgra, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv4, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv4, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv4, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv4, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgb, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgba, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgb, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgba, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgr, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgra, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgr, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgra, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls4, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls4, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls4, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls4, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgb, 3, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgba, 3, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgb, 4, 3, 2)
+    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgba, 4, 4, 2)
+    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgr, 3, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgra, 3, 4, 0)
+    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgr, 4, 3, 0)
+    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgra, 4, 4, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(rgb_to_lab, 3, 3, true, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(rgba_to_lab, 4, 3, true, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(rgb_to_lab4, 3, 4, true, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(rgba_to_lab4, 4, 4, true, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(bgr_to_lab, 3, 3, true, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(bgra_to_lab, 4, 3, true, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(bgr_to_lab4, 3, 4, true, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(bgra_to_lab4, 4, 4, true, 0)
+
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lrgb_to_lab, 3, 3, false, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lrgba_to_lab, 4, 3, false, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lrgb_to_lab4, 3, 4, false, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lrgba_to_lab4, 4, 4, false, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lbgr_to_lab, 3, 3, false, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lbgra_to_lab, 4, 3, false, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lbgr_to_lab4, 3, 4, false, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lbgra_to_lab4, 4, 4, false, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_rgb, 3, 3, true, 2)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_rgb, 4, 3, true, 2)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_rgba, 3, 4, true, 2)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_rgba, 4, 4, true, 2)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_bgr, 3, 3, true, 0)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_bgr, 4, 3, true, 0)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_bgra, 3, 4, true, 0)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_bgra, 4, 4, true, 0)
+
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_lrgb, 3, 3, false, 2)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_lrgb, 4, 3, false, 2)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_lrgba, 3, 4, false, 2)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_lrgba, 4, 4, false, 2)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_lbgr, 3, 3, false, 0)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_lbgr, 4, 3, false, 0)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_lbgra, 3, 4, false, 0)
+    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_lbgra, 4, 4, false, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(rgb_to_luv, 3, 3, true, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(rgba_to_luv, 4, 3, true, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(rgb_to_luv4, 3, 4, true, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(rgba_to_luv4, 4, 4, true, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(bgr_to_luv, 3, 3, true, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(bgra_to_luv, 4, 3, true, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(bgr_to_luv4, 3, 4, true, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(bgra_to_luv4, 4, 4, true, 0)
+
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lrgb_to_luv, 3, 3, false, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lrgba_to_luv, 4, 3, false, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lrgb_to_luv4, 3, 4, false, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lrgba_to_luv4, 4, 4, false, 2)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lbgr_to_luv, 3, 3, false, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lbgra_to_luv, 4, 3, false, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lbgr_to_luv4, 3, 4, false, 0)
+    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lbgra_to_luv4, 4, 4, false, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS
+
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_rgb, 3, 3, true, 2)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_rgb, 4, 3, true, 2)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_rgba, 3, 4, true, 2)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_rgba, 4, 4, true, 2)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_bgr, 3, 3, true, 0)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_bgr, 4, 3, true, 0)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_bgra, 3, 4, true, 0)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_bgra, 4, 4, true, 0)
+
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_lrgb, 3, 3, false, 2)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_lrgb, 4, 3, false, 2)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_lrgba, 3, 4, false, 2)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_lrgba, 4, 4, false, 2)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_lbgr, 3, 3, false, 0)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_lbgr, 4, 3, false, 0)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_lbgra, 3, 4, false, 0)
+    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_lbgra, 4, 4, false, 0)
+
+    #undef OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_BORDER_INTERPOLATE_HPP__

3rdparty/include/opencv2/core/cuda/common.hpp

@@ -0,0 +1,109 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_COMMON_HPP__
+#define __OPENCV_CUDA_COMMON_HPP__
+
+#include <cuda_runtime.h>
+#include "opencv2/core/cuda_types.hpp"
+#include "opencv2/core/cvdef.h"
+#include "opencv2/core/base.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+#ifndef CV_PI_F
+    #ifndef CV_PI
+        #define CV_PI_F 3.14159265f
+    #else
+        #define CV_PI_F ((float)CV_PI)
+    #endif
+#endif
+
+namespace cv { namespace cuda {
+    static inline void checkCudaError(cudaError_t err, const char* file, const int line, const char* func)
+    {
+        if (cudaSuccess != err)
+            cv::error(cv::Error::GpuApiCallError, cudaGetErrorString(err), func, file, line);
+    }
+}}
+
+#ifndef cudaSafeCall
+    #define cudaSafeCall(expr)  cv::cuda::checkCudaError(expr, __FILE__, __LINE__, CV_Func)
+#endif
+
+namespace cv { namespace cuda
+{
+    template <typename T> static inline bool isAligned(const T* ptr, size_t size)
+    {
+        return reinterpret_cast<size_t>(ptr) % size == 0;
+    }
+
+    static inline bool isAligned(size_t step, size_t size)
+    {
+        return step % size == 0;
+    }
+}}
+
+namespace cv { namespace cuda
+{
+    namespace device
+    {
+        __host__ __device__ __forceinline__ int divUp(int total, int grain)
+        {
+            return (total + grain - 1) / grain;
+        }
+
+        template<class T> inline void bindTexture(const textureReference* tex, const PtrStepSz<T>& img)
+        {
+            cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
+            cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
+        }
+    }
+}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_COMMON_HPP__

3rdparty/include/opencv2/core/cuda/datamov_utils.hpp

@@ -0,0 +1,113 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_DATAMOV_UTILS_HPP__
+#define __OPENCV_CUDA_DATAMOV_UTILS_HPP__
+
+#include "common.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 200
+
+        // for Fermi memory space is detected automatically
+        template <typename T> struct ForceGlob
+        {
+            __device__ __forceinline__ static void Load(const T* ptr, int offset, T& val)  { val = ptr[offset];  }
+        };
+
+    #else // __CUDA_ARCH__ >= 200
+
+        #if defined(_WIN64) || defined(__LP64__)
+            // 64-bit register modifier for inlined asm
+            #define OPENCV_CUDA_ASM_PTR "l"
+        #else
+            // 32-bit register modifier for inlined asm
+            #define OPENCV_CUDA_ASM_PTR "r"
+        #endif
+
+        template<class T> struct ForceGlob;
+
+        #define OPENCV_CUDA_DEFINE_FORCE_GLOB(base_type, ptx_type, reg_mod) \
+            template <> struct ForceGlob<base_type> \
+            { \
+                __device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
+                { \
+                    asm("ld.global."#ptx_type" %0, [%1];" : "="#reg_mod(val) : OPENCV_CUDA_ASM_PTR(ptr + offset)); \
+                } \
+            };
+
+        #define OPENCV_CUDA_DEFINE_FORCE_GLOB_B(base_type, ptx_type) \
+            template <> struct ForceGlob<base_type> \
+            { \
+                __device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
+                { \
+                    asm("ld.global."#ptx_type" %0, [%1];" : "=r"(*reinterpret_cast<uint*>(&val)) : OPENCV_CUDA_ASM_PTR(ptr + offset)); \
+                } \
+            };
+
+            OPENCV_CUDA_DEFINE_FORCE_GLOB_B(uchar,  u8)
+            OPENCV_CUDA_DEFINE_FORCE_GLOB_B(schar,  s8)
+            OPENCV_CUDA_DEFINE_FORCE_GLOB_B(char,   b8)
+            OPENCV_CUDA_DEFINE_FORCE_GLOB  (ushort, u16, h)
+            OPENCV_CUDA_DEFINE_FORCE_GLOB  (short,  s16, h)
+            OPENCV_CUDA_DEFINE_FORCE_GLOB  (uint,   u32, r)
+            OPENCV_CUDA_DEFINE_FORCE_GLOB  (int,    s32, r)
+            OPENCV_CUDA_DEFINE_FORCE_GLOB  (float,  f32, f)
+            OPENCV_CUDA_DEFINE_FORCE_GLOB  (double, f64, d)
+
+        #undef OPENCV_CUDA_DEFINE_FORCE_GLOB
+        #undef OPENCV_CUDA_DEFINE_FORCE_GLOB_B
+        #undef OPENCV_CUDA_ASM_PTR
+
+    #endif // __CUDA_ARCH__ >= 200
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_DATAMOV_UTILS_HPP__

3rdparty/include/opencv2/core/cuda/detail/reduce.hpp

@@ -0,0 +1,365 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_REDUCE_DETAIL_HPP__
+#define __OPENCV_CUDA_REDUCE_DETAIL_HPP__
+
+#include <thrust/tuple.h>
+#include "../warp.hpp"
+#include "../warp_shuffle.hpp"
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    namespace reduce_detail
+    {
+        template <typename T> struct GetType;
+        template <typename T> struct GetType<T*>
+        {
+            typedef T type;
+        };
+        template <typename T> struct GetType<volatile T*>
+        {
+            typedef T type;
+        };
+        template <typename T> struct GetType<T&>
+        {
+            typedef T type;
+        };
+
+        template <unsigned int I, unsigned int N>
+        struct For
+        {
+            template <class PointerTuple, class ValTuple>
+            static __device__ void loadToSmem(const PointerTuple& smem, const ValTuple& val, unsigned int tid)
+            {
+                thrust::get<I>(smem)[tid] = thrust::get<I>(val);
+
+                For<I + 1, N>::loadToSmem(smem, val, tid);
+            }
+            template <class PointerTuple, class ValTuple>
+            static __device__ void loadFromSmem(const PointerTuple& smem, const ValTuple& val, unsigned int tid)
+            {
+                thrust::get<I>(val) = thrust::get<I>(smem)[tid];
+
+                For<I + 1, N>::loadFromSmem(smem, val, tid);
+            }
+
+            template <class PointerTuple, class ValTuple, class OpTuple>
+            static __device__ void merge(const PointerTuple& smem, const ValTuple& val, unsigned int tid, unsigned int delta, const OpTuple& op)
+            {
+                typename GetType<typename thrust::tuple_element<I, PointerTuple>::type>::type reg = thrust::get<I>(smem)[tid + delta];
+                thrust::get<I>(smem)[tid] = thrust::get<I>(val) = thrust::get<I>(op)(thrust::get<I>(val), reg);
+
+                For<I + 1, N>::merge(smem, val, tid, delta, op);
+            }
+            template <class ValTuple, class OpTuple>
+            static __device__ void mergeShfl(const ValTuple& val, unsigned int delta, unsigned int width, const OpTuple& op)
+            {
+                typename GetType<typename thrust::tuple_element<I, ValTuple>::type>::type reg = shfl_down(thrust::get<I>(val), delta, width);
+                thrust::get<I>(val) = thrust::get<I>(op)(thrust::get<I>(val), reg);
+
+                For<I + 1, N>::mergeShfl(val, delta, width, op);
+            }
+        };
+        template <unsigned int N>
+        struct For<N, N>
+        {
+            template <class PointerTuple, class ValTuple>
+            static __device__ void loadToSmem(const PointerTuple&, const ValTuple&, unsigned int)
+            {
+            }
+            template <class PointerTuple, class ValTuple>
+            static __device__ void loadFromSmem(const PointerTuple&, const ValTuple&, unsigned int)
+            {
+            }
+
+            template <class PointerTuple, class ValTuple, class OpTuple>
+            static __device__ void merge(const PointerTuple&, const ValTuple&, unsigned int, unsigned int, const OpTuple&)
+            {
+            }
+            template <class ValTuple, class OpTuple>
+            static __device__ void mergeShfl(const ValTuple&, unsigned int, unsigned int, const OpTuple&)
+            {
+            }
+        };
+
+        template <typename T>
+        __device__ __forceinline__ void loadToSmem(volatile T* smem, T& val, unsigned int tid)
+        {
+            smem[tid] = val;
+        }
+        template <typename T>
+        __device__ __forceinline__ void loadFromSmem(volatile T* smem, T& val, unsigned int tid)
+        {
+            val = smem[tid];
+        }
+        template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
+                  typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9>
+        __device__ __forceinline__ void loadToSmem(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
+                                                       const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
+                                                       unsigned int tid)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9> >::value>::loadToSmem(smem, val, tid);
+        }
+        template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
+                  typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9>
+        __device__ __forceinline__ void loadFromSmem(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
+                                                         const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
+                                                         unsigned int tid)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9> >::value>::loadFromSmem(smem, val, tid);
+        }
+
+        template <typename T, class Op>
+        __device__ __forceinline__ void merge(volatile T* smem, T& val, unsigned int tid, unsigned int delta, const Op& op)
+        {
+            T reg = smem[tid + delta];
+            smem[tid] = val = op(val, reg);
+        }
+        template <typename T, class Op>
+        __device__ __forceinline__ void mergeShfl(T& val, unsigned int delta, unsigned int width, const Op& op)
+        {
+            T reg = shfl_down(val, delta, width);
+            val = op(val, reg);
+        }
+        template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
+                  typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9,
+                  class Op0, class Op1, class Op2, class Op3, class Op4, class Op5, class Op6, class Op7, class Op8, class Op9>
+        __device__ __forceinline__ void merge(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
+                                              const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
+                                              unsigned int tid,
+                                              unsigned int delta,
+                                              const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>& op)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9> >::value>::merge(smem, val, tid, delta, op);
+        }
+        template <typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9,
+                  class Op0, class Op1, class Op2, class Op3, class Op4, class Op5, class Op6, class Op7, class Op8, class Op9>
+        __device__ __forceinline__ void mergeShfl(const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
+                                                  unsigned int delta,
+                                                  unsigned int width,
+                                                  const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>& op)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9> >::value>::mergeShfl(val, delta, width, op);
+        }
+
+        template <unsigned int N> struct Generic
+        {
+            template <typename Pointer, typename Reference, class Op>
+            static __device__ void reduce(Pointer smem, Reference val, unsigned int tid, Op op)
+            {
+                loadToSmem(smem, val, tid);
+                if (N >= 32)
+                    __syncthreads();
+
+                if (N >= 2048)
+                {
+                    if (tid < 1024)
+                        merge(smem, val, tid, 1024, op);
+
+                    __syncthreads();
+                }
+                if (N >= 1024)
+                {
+                    if (tid < 512)
+                        merge(smem, val, tid, 512, op);
+
+                    __syncthreads();
+                }
+                if (N >= 512)
+                {
+                    if (tid < 256)
+                        merge(smem, val, tid, 256, op);
+
+                    __syncthreads();
+                }
+                if (N >= 256)
+                {
+                    if (tid < 128)
+                        merge(smem, val, tid, 128, op);
+
+                    __syncthreads();
+                }
+                if (N >= 128)
+                {
+                    if (tid < 64)
+                        merge(smem, val, tid, 64, op);
+
+                    __syncthreads();
+                }
+                if (N >= 64)
+                {
+                    if (tid < 32)
+                        merge(smem, val, tid, 32, op);
+                }
+
+                if (tid < 16)
+                {
+                    merge(smem, val, tid, 16, op);
+                    merge(smem, val, tid, 8, op);
+                    merge(smem, val, tid, 4, op);
+                    merge(smem, val, tid, 2, op);
+                    merge(smem, val, tid, 1, op);
+                }
+            }
+        };
+
+        template <unsigned int I, typename Pointer, typename Reference, class Op>
+        struct Unroll
+        {
+            static __device__ void loopShfl(Reference val, Op op, unsigned int N)
+            {
+                mergeShfl(val, I, N, op);
+                Unroll<I / 2, Pointer, Reference, Op>::loopShfl(val, op, N);
+            }
+            static __device__ void loop(Pointer smem, Reference val, unsigned int tid, Op op)
+            {
+                merge(smem, val, tid, I, op);
+                Unroll<I / 2, Pointer, Reference, Op>::loop(smem, val, tid, op);
+            }
+        };
+        template <typename Pointer, typename Reference, class Op>
+        struct Unroll<0, Pointer, Reference, Op>
+        {
+            static __device__ void loopShfl(Reference, Op, unsigned int)
+            {
+            }
+            static __device__ void loop(Pointer, Reference, unsigned int, Op)
+            {
+            }
+        };
+
+        template <unsigned int N> struct WarpOptimized
+        {
+            template <typename Pointer, typename Reference, class Op>
+            static __device__ void reduce(Pointer smem, Reference val, unsigned int tid, Op op)
+            {
+            #if __CUDA_ARCH__ >= 300
+                (void) smem;
+                (void) tid;
+
+                Unroll<N / 2, Pointer, Reference, Op>::loopShfl(val, op, N);
+            #else
+                loadToSmem(smem, val, tid);
+
+                if (tid < N / 2)
+                    Unroll<N / 2, Pointer, Reference, Op>::loop(smem, val, tid, op);
+            #endif
+            }
+        };
+
+        template <unsigned int N> struct GenericOptimized32
+        {
+            enum { M = N / 32 };
+
+            template <typename Pointer, typename Reference, class Op>
+            static __device__ void reduce(Pointer smem, Reference val, unsigned int tid, Op op)
+            {
+                const unsigned int laneId = Warp::laneId();
+
+            #if __CUDA_ARCH__ >= 300
+                Unroll<16, Pointer, Reference, Op>::loopShfl(val, op, warpSize);
+
+                if (laneId == 0)
+                    loadToSmem(smem, val, tid / 32);
+            #else
+                loadToSmem(smem, val, tid);
+
+                if (laneId < 16)
+                    Unroll<16, Pointer, Reference, Op>::loop(smem, val, tid, op);
+
+                __syncthreads();
+
+                if (laneId == 0)
+                    loadToSmem(smem, val, tid / 32);
+            #endif
+
+                __syncthreads();
+
+                loadFromSmem(smem, val, tid);
+
+                if (tid < 32)
+                {
+                #if __CUDA_ARCH__ >= 300
+                    Unroll<M / 2, Pointer, Reference, Op>::loopShfl(val, op, M);
+                #else
+                    Unroll<M / 2, Pointer, Reference, Op>::loop(smem, val, tid, op);
+                #endif
+                }
+            }
+        };
+
+        template <bool val, class T1, class T2> struct StaticIf;
+        template <class T1, class T2> struct StaticIf<true, T1, T2>
+        {
+            typedef T1 type;
+        };
+        template <class T1, class T2> struct StaticIf<false, T1, T2>
+        {
+            typedef T2 type;
+        };
+
+        template <unsigned int N> struct IsPowerOf2
+        {
+            enum { value = ((N != 0) && !(N & (N - 1))) };
+        };
+
+        template <unsigned int N> struct Dispatcher
+        {
+            typedef typename StaticIf<
+                (N <= 32) && IsPowerOf2<N>::value,
+                WarpOptimized<N>,
+                typename StaticIf<
+                    (N <= 1024) && IsPowerOf2<N>::value,
+                    GenericOptimized32<N>,
+                    Generic<N>
+                >::type
+            >::type reductor;
+        };
+    }
+}}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_REDUCE_DETAIL_HPP__

3rdparty/include/opencv2/core/cuda/detail/reduce_key_val.hpp

@@ -0,0 +1,502 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_PRED_VAL_REDUCE_DETAIL_HPP__
+#define __OPENCV_CUDA_PRED_VAL_REDUCE_DETAIL_HPP__
+
+#include <thrust/tuple.h>
+#include "../warp.hpp"
+#include "../warp_shuffle.hpp"
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    namespace reduce_key_val_detail
+    {
+        template <typename T> struct GetType;
+        template <typename T> struct GetType<T*>
+        {
+            typedef T type;
+        };
+        template <typename T> struct GetType<volatile T*>
+        {
+            typedef T type;
+        };
+        template <typename T> struct GetType<T&>
+        {
+            typedef T type;
+        };
+
+        template <unsigned int I, unsigned int N>
+        struct For
+        {
+            template <class PointerTuple, class ReferenceTuple>
+            static __device__ void loadToSmem(const PointerTuple& smem, const ReferenceTuple& data, unsigned int tid)
+            {
+                thrust::get<I>(smem)[tid] = thrust::get<I>(data);
+
+                For<I + 1, N>::loadToSmem(smem, data, tid);
+            }
+            template <class PointerTuple, class ReferenceTuple>
+            static __device__ void loadFromSmem(const PointerTuple& smem, const ReferenceTuple& data, unsigned int tid)
+            {
+                thrust::get<I>(data) = thrust::get<I>(smem)[tid];
+
+                For<I + 1, N>::loadFromSmem(smem, data, tid);
+            }
+
+            template <class ReferenceTuple>
+            static __device__ void copyShfl(const ReferenceTuple& val, unsigned int delta, int width)
+            {
+                thrust::get<I>(val) = shfl_down(thrust::get<I>(val), delta, width);
+
+                For<I + 1, N>::copyShfl(val, delta, width);
+            }
+            template <class PointerTuple, class ReferenceTuple>
+            static __device__ void copy(const PointerTuple& svals, const ReferenceTuple& val, unsigned int tid, unsigned int delta)
+            {
+                thrust::get<I>(svals)[tid] = thrust::get<I>(val) = thrust::get<I>(svals)[tid + delta];
+
+                For<I + 1, N>::copy(svals, val, tid, delta);
+            }
+
+            template <class KeyReferenceTuple, class ValReferenceTuple, class CmpTuple>
+            static __device__ void mergeShfl(const KeyReferenceTuple& key, const ValReferenceTuple& val, const CmpTuple& cmp, unsigned int delta, int width)
+            {
+                typename GetType<typename thrust::tuple_element<I, KeyReferenceTuple>::type>::type reg = shfl_down(thrust::get<I>(key), delta, width);
+
+                if (thrust::get<I>(cmp)(reg, thrust::get<I>(key)))
+                {
+                    thrust::get<I>(key) = reg;
+                    thrust::get<I>(val) = shfl_down(thrust::get<I>(val), delta, width);
+                }
+
+                For<I + 1, N>::mergeShfl(key, val, cmp, delta, width);
+            }
+            template <class KeyPointerTuple, class KeyReferenceTuple, class ValPointerTuple, class ValReferenceTuple, class CmpTuple>
+            static __device__ void merge(const KeyPointerTuple& skeys, const KeyReferenceTuple& key,
+                                         const ValPointerTuple& svals, const ValReferenceTuple& val,
+                                         const CmpTuple& cmp,
+                                         unsigned int tid, unsigned int delta)
+            {
+                typename GetType<typename thrust::tuple_element<I, KeyPointerTuple>::type>::type reg = thrust::get<I>(skeys)[tid + delta];
+
+                if (thrust::get<I>(cmp)(reg, thrust::get<I>(key)))
+                {
+                    thrust::get<I>(skeys)[tid] = thrust::get<I>(key) = reg;
+                    thrust::get<I>(svals)[tid] = thrust::get<I>(val) = thrust::get<I>(svals)[tid + delta];
+                }
+
+                For<I + 1, N>::merge(skeys, key, svals, val, cmp, tid, delta);
+            }
+        };
+        template <unsigned int N>
+        struct For<N, N>
+        {
+            template <class PointerTuple, class ReferenceTuple>
+            static __device__ void loadToSmem(const PointerTuple&, const ReferenceTuple&, unsigned int)
+            {
+            }
+            template <class PointerTuple, class ReferenceTuple>
+            static __device__ void loadFromSmem(const PointerTuple&, const ReferenceTuple&, unsigned int)
+            {
+            }
+
+            template <class ReferenceTuple>
+            static __device__ void copyShfl(const ReferenceTuple&, unsigned int, int)
+            {
+            }
+            template <class PointerTuple, class ReferenceTuple>
+            static __device__ void copy(const PointerTuple&, const ReferenceTuple&, unsigned int, unsigned int)
+            {
+            }
+
+            template <class KeyReferenceTuple, class ValReferenceTuple, class CmpTuple>
+            static __device__ void mergeShfl(const KeyReferenceTuple&, const ValReferenceTuple&, const CmpTuple&, unsigned int, int)
+            {
+            }
+            template <class KeyPointerTuple, class KeyReferenceTuple, class ValPointerTuple, class ValReferenceTuple, class CmpTuple>
+            static __device__ void merge(const KeyPointerTuple&, const KeyReferenceTuple&,
+                                         const ValPointerTuple&, const ValReferenceTuple&,
+                                         const CmpTuple&,
+                                         unsigned int, unsigned int)
+            {
+            }
+        };
+
+        //////////////////////////////////////////////////////
+        // loadToSmem
+
+        template <typename T>
+        __device__ __forceinline__ void loadToSmem(volatile T* smem, T& data, unsigned int tid)
+        {
+            smem[tid] = data;
+        }
+        template <typename T>
+        __device__ __forceinline__ void loadFromSmem(volatile T* smem, T& data, unsigned int tid)
+        {
+            data = smem[tid];
+        }
+        template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
+                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
+        __device__ __forceinline__ void loadToSmem(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& smem,
+                                                   const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& data,
+                                                   unsigned int tid)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::loadToSmem(smem, data, tid);
+        }
+        template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
+                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
+        __device__ __forceinline__ void loadFromSmem(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& smem,
+                                                     const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& data,
+                                                     unsigned int tid)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::loadFromSmem(smem, data, tid);
+        }
+
+        //////////////////////////////////////////////////////
+        // copyVals
+
+        template <typename V>
+        __device__ __forceinline__ void copyValsShfl(V& val, unsigned int delta, int width)
+        {
+            val = shfl_down(val, delta, width);
+        }
+        template <typename V>
+        __device__ __forceinline__ void copyVals(volatile V* svals, V& val, unsigned int tid, unsigned int delta)
+        {
+            svals[tid] = val = svals[tid + delta];
+        }
+        template <typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
+        __device__ __forceinline__ void copyValsShfl(const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
+                                                     unsigned int delta,
+                                                     int width)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9> >::value>::copyShfl(val, delta, width);
+        }
+        template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
+                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
+        __device__ __forceinline__ void copyVals(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
+                                                 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
+                                                 unsigned int tid, unsigned int delta)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::copy(svals, val, tid, delta);
+        }
+
+        //////////////////////////////////////////////////////
+        // merge
+
+        template <typename K, typename V, class Cmp>
+        __device__ __forceinline__ void mergeShfl(K& key, V& val, const Cmp& cmp, unsigned int delta, int width)
+        {
+            K reg = shfl_down(key, delta, width);
+
+            if (cmp(reg, key))
+            {
+                key = reg;
+                copyValsShfl(val, delta, width);
+            }
+        }
+        template <typename K, typename V, class Cmp>
+        __device__ __forceinline__ void merge(volatile K* skeys, K& key, volatile V* svals, V& val, const Cmp& cmp, unsigned int tid, unsigned int delta)
+        {
+            K reg = skeys[tid + delta];
+
+            if (cmp(reg, key))
+            {
+                skeys[tid] = key = reg;
+                copyVals(svals, val, tid, delta);
+            }
+        }
+        template <typename K,
+                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
+                  class Cmp>
+        __device__ __forceinline__ void mergeShfl(K& key,
+                                                  const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
+                                                  const Cmp& cmp,
+                                                  unsigned int delta, int width)
+        {
+            K reg = shfl_down(key, delta, width);
+
+            if (cmp(reg, key))
+            {
+                key = reg;
+                copyValsShfl(val, delta, width);
+            }
+        }
+        template <typename K,
+                  typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
+                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
+                  class Cmp>
+        __device__ __forceinline__ void merge(volatile K* skeys, K& key,
+                                              const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
+                                              const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
+                                              const Cmp& cmp, unsigned int tid, unsigned int delta)
+        {
+            K reg = skeys[tid + delta];
+
+            if (cmp(reg, key))
+            {
+                skeys[tid] = key = reg;
+                copyVals(svals, val, tid, delta);
+            }
+        }
+        template <typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9,
+                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
+                  class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9>
+        __device__ __forceinline__ void mergeShfl(const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key,
+                                                  const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
+                                                  const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp,
+                                                  unsigned int delta, int width)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9> >::value>::mergeShfl(key, val, cmp, delta, width);
+        }
+        template <typename KP0, typename KP1, typename KP2, typename KP3, typename KP4, typename KP5, typename KP6, typename KP7, typename KP8, typename KP9,
+                  typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9,
+                  typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
+                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
+                  class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9>
+        __device__ __forceinline__ void merge(const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>& skeys,
+                                              const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key,
+                                              const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
+                                              const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
+                                              const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp,
+                                              unsigned int tid, unsigned int delta)
+        {
+            For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::merge(skeys, key, svals, val, cmp, tid, delta);
+        }
+
+        //////////////////////////////////////////////////////
+        // Generic
+
+        template <unsigned int N> struct Generic
+        {
+            template <class KP, class KR, class VP, class VR, class Cmp>
+            static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
+            {
+                loadToSmem(skeys, key, tid);
+                loadValsToSmem(svals, val, tid);
+                if (N >= 32)
+                    __syncthreads();
+
+                if (N >= 2048)
+                {
+                    if (tid < 1024)
+                        merge(skeys, key, svals, val, cmp, tid, 1024);
+
+                    __syncthreads();
+                }
+                if (N >= 1024)
+                {
+                    if (tid < 512)
+                        merge(skeys, key, svals, val, cmp, tid, 512);
+
+                    __syncthreads();
+                }
+                if (N >= 512)
+                {
+                    if (tid < 256)
+                        merge(skeys, key, svals, val, cmp, tid, 256);
+
+                    __syncthreads();
+                }
+                if (N >= 256)
+                {
+                    if (tid < 128)
+                        merge(skeys, key, svals, val, cmp, tid, 128);
+
+                    __syncthreads();
+                }
+                if (N >= 128)
+                {
+                    if (tid < 64)
+                        merge(skeys, key, svals, val, cmp, tid, 64);
+
+                    __syncthreads();
+                }
+                if (N >= 64)
+                {
+                    if (tid < 32)
+                        merge(skeys, key, svals, val, cmp, tid, 32);
+                }
+
+                if (tid < 16)
+                {
+                    merge(skeys, key, svals, val, cmp, tid, 16);
+                    merge(skeys, key, svals, val, cmp, tid, 8);
+                    merge(skeys, key, svals, val, cmp, tid, 4);
+                    merge(skeys, key, svals, val, cmp, tid, 2);
+                    merge(skeys, key, svals, val, cmp, tid, 1);
+                }
+            }
+        };
+
+        template <unsigned int I, class KP, class KR, class VP, class VR, class Cmp>
+        struct Unroll
+        {
+            static __device__ void loopShfl(KR key, VR val, Cmp cmp, unsigned int N)
+            {
+                mergeShfl(key, val, cmp, I, N);
+                Unroll<I / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, N);
+            }
+            static __device__ void loop(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
+            {
+                merge(skeys, key, svals, val, cmp, tid, I);
+                Unroll<I / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
+            }
+        };
+        template <class KP, class KR, class VP, class VR, class Cmp>
+        struct Unroll<0, KP, KR, VP, VR, Cmp>
+        {
+            static __device__ void loopShfl(KR, VR, Cmp, unsigned int)
+            {
+            }
+            static __device__ void loop(KP, KR, VP, VR, unsigned int, Cmp)
+            {
+            }
+        };
+
+        template <unsigned int N> struct WarpOptimized
+        {
+            template <class KP, class KR, class VP, class VR, class Cmp>
+            static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
+            {
+            #if 0 // __CUDA_ARCH__ >= 300
+                (void) skeys;
+                (void) svals;
+                (void) tid;
+
+                Unroll<N / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, N);
+            #else
+                loadToSmem(skeys, key, tid);
+                loadToSmem(svals, val, tid);
+
+                if (tid < N / 2)
+                    Unroll<N / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
+            #endif
+            }
+        };
+
+        template <unsigned int N> struct GenericOptimized32
+        {
+            enum { M = N / 32 };
+
+            template <class KP, class KR, class VP, class VR, class Cmp>
+            static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
+            {
+                const unsigned int laneId = Warp::laneId();
+
+            #if 0 // __CUDA_ARCH__ >= 300
+                Unroll<16, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, warpSize);
+
+                if (laneId == 0)
+                {
+                    loadToSmem(skeys, key, tid / 32);
+                    loadToSmem(svals, val, tid / 32);
+                }
+            #else
+                loadToSmem(skeys, key, tid);
+                loadToSmem(svals, val, tid);
+
+                if (laneId < 16)
+                    Unroll<16, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
+
+                __syncthreads();
+
+                if (laneId == 0)
+                {
+                    loadToSmem(skeys, key, tid / 32);
+                    loadToSmem(svals, val, tid / 32);
+                }
+            #endif
+
+                __syncthreads();
+
+                loadFromSmem(skeys, key, tid);
+
+                if (tid < 32)
+                {
+                #if 0 // __CUDA_ARCH__ >= 300
+                    loadFromSmem(svals, val, tid);
+
+                    Unroll<M / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, M);
+                #else
+                    Unroll<M / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
+                #endif
+                }
+            }
+        };
+
+        template <bool val, class T1, class T2> struct StaticIf;
+        template <class T1, class T2> struct StaticIf<true, T1, T2>
+        {
+            typedef T1 type;
+        };
+        template <class T1, class T2> struct StaticIf<false, T1, T2>
+        {
+            typedef T2 type;
+        };
+
+        template <unsigned int N> struct IsPowerOf2
+        {
+            enum { value = ((N != 0) && !(N & (N - 1))) };
+        };
+
+        template <unsigned int N> struct Dispatcher
+        {
+            typedef typename StaticIf<
+                (N <= 32) && IsPowerOf2<N>::value,
+                WarpOptimized<N>,
+                typename StaticIf<
+                    (N <= 1024) && IsPowerOf2<N>::value,
+                    GenericOptimized32<N>,
+                    Generic<N>
+                >::type
+            >::type reductor;
+        };
+    }
+}}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_PRED_VAL_REDUCE_DETAIL_HPP__

3rdparty/include/opencv2/core/cuda/detail/transform_detail.hpp

@@ -0,0 +1,399 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_TRANSFORM_DETAIL_HPP__
+#define __OPENCV_CUDA_TRANSFORM_DETAIL_HPP__
+
+#include "../common.hpp"
+#include "../vec_traits.hpp"
+#include "../functional.hpp"
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    namespace transform_detail
+    {
+        //! Read Write Traits
+
+        template <typename T, typename D, int shift> struct UnaryReadWriteTraits
+        {
+            typedef typename TypeVec<T, shift>::vec_type read_type;
+            typedef typename TypeVec<D, shift>::vec_type write_type;
+        };
+
+        template <typename T1, typename T2, typename D, int shift> struct BinaryReadWriteTraits
+        {
+            typedef typename TypeVec<T1, shift>::vec_type read_type1;
+            typedef typename TypeVec<T2, shift>::vec_type read_type2;
+            typedef typename TypeVec<D, shift>::vec_type write_type;
+        };
+
+        //! Transform kernels
+
+        template <int shift> struct OpUnroller;
+        template <> struct OpUnroller<1>
+        {
+            template <typename T, typename D, typename UnOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.x = op(src.x);
+            }
+
+            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.x = op(src1.x, src2.x);
+            }
+        };
+        template <> struct OpUnroller<2>
+        {
+            template <typename T, typename D, typename UnOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.x = op(src.x);
+                if (mask(y, x_shifted + 1))
+                    dst.y = op(src.y);
+            }
+
+            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.x = op(src1.x, src2.x);
+                if (mask(y, x_shifted + 1))
+                    dst.y = op(src1.y, src2.y);
+            }
+        };
+        template <> struct OpUnroller<3>
+        {
+            template <typename T, typename D, typename UnOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.x = op(src.x);
+                if (mask(y, x_shifted + 1))
+                    dst.y = op(src.y);
+                if (mask(y, x_shifted + 2))
+                    dst.z = op(src.z);
+            }
+
+            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.x = op(src1.x, src2.x);
+                if (mask(y, x_shifted + 1))
+                    dst.y = op(src1.y, src2.y);
+                if (mask(y, x_shifted + 2))
+                    dst.z = op(src1.z, src2.z);
+            }
+        };
+        template <> struct OpUnroller<4>
+        {
+            template <typename T, typename D, typename UnOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.x = op(src.x);
+                if (mask(y, x_shifted + 1))
+                    dst.y = op(src.y);
+                if (mask(y, x_shifted + 2))
+                    dst.z = op(src.z);
+                if (mask(y, x_shifted + 3))
+                    dst.w = op(src.w);
+            }
+
+            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.x = op(src1.x, src2.x);
+                if (mask(y, x_shifted + 1))
+                    dst.y = op(src1.y, src2.y);
+                if (mask(y, x_shifted + 2))
+                    dst.z = op(src1.z, src2.z);
+                if (mask(y, x_shifted + 3))
+                    dst.w = op(src1.w, src2.w);
+            }
+        };
+        template <> struct OpUnroller<8>
+        {
+            template <typename T, typename D, typename UnOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.a0 = op(src.a0);
+                if (mask(y, x_shifted + 1))
+                    dst.a1 = op(src.a1);
+                if (mask(y, x_shifted + 2))
+                    dst.a2 = op(src.a2);
+                if (mask(y, x_shifted + 3))
+                    dst.a3 = op(src.a3);
+                if (mask(y, x_shifted + 4))
+                    dst.a4 = op(src.a4);
+                if (mask(y, x_shifted + 5))
+                    dst.a5 = op(src.a5);
+                if (mask(y, x_shifted + 6))
+                    dst.a6 = op(src.a6);
+                if (mask(y, x_shifted + 7))
+                    dst.a7 = op(src.a7);
+            }
+
+            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
+            {
+                if (mask(y, x_shifted))
+                    dst.a0 = op(src1.a0, src2.a0);
+                if (mask(y, x_shifted + 1))
+                    dst.a1 = op(src1.a1, src2.a1);
+                if (mask(y, x_shifted + 2))
+                    dst.a2 = op(src1.a2, src2.a2);
+                if (mask(y, x_shifted + 3))
+                    dst.a3 = op(src1.a3, src2.a3);
+                if (mask(y, x_shifted + 4))
+                    dst.a4 = op(src1.a4, src2.a4);
+                if (mask(y, x_shifted + 5))
+                    dst.a5 = op(src1.a5, src2.a5);
+                if (mask(y, x_shifted + 6))
+                    dst.a6 = op(src1.a6, src2.a6);
+                if (mask(y, x_shifted + 7))
+                    dst.a7 = op(src1.a7, src2.a7);
+            }
+        };
+
+        template <typename T, typename D, typename UnOp, typename Mask>
+        static __global__ void transformSmart(const PtrStepSz<T> src_, PtrStep<D> dst_, const Mask mask, const UnOp op)
+        {
+            typedef TransformFunctorTraits<UnOp> ft;
+            typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::read_type read_type;
+            typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::write_type write_type;
+
+            const int x = threadIdx.x + blockIdx.x * blockDim.x;
+            const int y = threadIdx.y + blockIdx.y * blockDim.y;
+            const int x_shifted = x * ft::smart_shift;
+
+            if (y < src_.rows)
+            {
+                const T* src = src_.ptr(y);
+                D* dst = dst_.ptr(y);
+
+                if (x_shifted + ft::smart_shift - 1 < src_.cols)
+                {
+                    const read_type src_n_el = ((const read_type*)src)[x];
+                    write_type dst_n_el = ((const write_type*)dst)[x];
+
+                    OpUnroller<ft::smart_shift>::unroll(src_n_el, dst_n_el, mask, op, x_shifted, y);
+
+                    ((write_type*)dst)[x] = dst_n_el;
+                }
+                else
+                {
+                    for (int real_x = x_shifted; real_x < src_.cols; ++real_x)
+                    {
+                        if (mask(y, real_x))
+                            dst[real_x] = op(src[real_x]);
+                    }
+                }
+            }
+        }
+
+        template <typename T, typename D, typename UnOp, typename Mask>
+        __global__ static void transformSimple(const PtrStepSz<T> src, PtrStep<D> dst, const Mask mask, const UnOp op)
+        {
+            const int x = blockDim.x * blockIdx.x + threadIdx.x;
+            const int y = blockDim.y * blockIdx.y + threadIdx.y;
+
+            if (x < src.cols && y < src.rows && mask(y, x))
+            {
+                dst.ptr(y)[x] = op(src.ptr(y)[x]);
+            }
+        }
+
+        template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+        static __global__ void transformSmart(const PtrStepSz<T1> src1_, const PtrStep<T2> src2_, PtrStep<D> dst_,
+            const Mask mask, const BinOp op)
+        {
+            typedef TransformFunctorTraits<BinOp> ft;
+            typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type1 read_type1;
+            typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type2 read_type2;
+            typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::write_type write_type;
+
+            const int x = threadIdx.x + blockIdx.x * blockDim.x;
+            const int y = threadIdx.y + blockIdx.y * blockDim.y;
+            const int x_shifted = x * ft::smart_shift;
+
+            if (y < src1_.rows)
+            {
+                const T1* src1 = src1_.ptr(y);
+                const T2* src2 = src2_.ptr(y);
+                D* dst = dst_.ptr(y);
+
+                if (x_shifted + ft::smart_shift - 1 < src1_.cols)
+                {
+                    const read_type1 src1_n_el = ((const read_type1*)src1)[x];
+                    const read_type2 src2_n_el = ((const read_type2*)src2)[x];
+                    write_type dst_n_el = ((const write_type*)dst)[x];
+
+                    OpUnroller<ft::smart_shift>::unroll(src1_n_el, src2_n_el, dst_n_el, mask, op, x_shifted, y);
+
+                    ((write_type*)dst)[x] = dst_n_el;
+                }
+                else
+                {
+                    for (int real_x = x_shifted; real_x < src1_.cols; ++real_x)
+                    {
+                        if (mask(y, real_x))
+                            dst[real_x] = op(src1[real_x], src2[real_x]);
+                    }
+                }
+            }
+        }
+
+        template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+        static __global__ void transformSimple(const PtrStepSz<T1> src1, const PtrStep<T2> src2, PtrStep<D> dst,
+            const Mask mask, const BinOp op)
+        {
+            const int x = blockDim.x * blockIdx.x + threadIdx.x;
+            const int y = blockDim.y * blockIdx.y + threadIdx.y;
+
+            if (x < src1.cols && y < src1.rows && mask(y, x))
+            {
+                const T1 src1_data = src1.ptr(y)[x];
+                const T2 src2_data = src2.ptr(y)[x];
+                dst.ptr(y)[x] = op(src1_data, src2_data);
+            }
+        }
+
+        template <bool UseSmart> struct TransformDispatcher;
+        template<> struct TransformDispatcher<false>
+        {
+            template <typename T, typename D, typename UnOp, typename Mask>
+            static void call(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, Mask mask, cudaStream_t stream)
+            {
+                typedef TransformFunctorTraits<UnOp> ft;
+
+                const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
+                const dim3 grid(divUp(src.cols, threads.x), divUp(src.rows, threads.y), 1);
+
+                transformSimple<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
+                cudaSafeCall( cudaGetLastError() );
+
+                if (stream == 0)
+                    cudaSafeCall( cudaDeviceSynchronize() );
+            }
+
+            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+            static void call(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, Mask mask, cudaStream_t stream)
+            {
+                typedef TransformFunctorTraits<BinOp> ft;
+
+                const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
+                const dim3 grid(divUp(src1.cols, threads.x), divUp(src1.rows, threads.y), 1);
+
+                transformSimple<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
+                cudaSafeCall( cudaGetLastError() );
+
+                if (stream == 0)
+                    cudaSafeCall( cudaDeviceSynchronize() );
+            }
+        };
+        template<> struct TransformDispatcher<true>
+        {
+            template <typename T, typename D, typename UnOp, typename Mask>
+            static void call(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, Mask mask, cudaStream_t stream)
+            {
+                typedef TransformFunctorTraits<UnOp> ft;
+
+                CV_StaticAssert(ft::smart_shift != 1, "");
+
+                if (!isAligned(src.data, ft::smart_shift * sizeof(T)) || !isAligned(src.step, ft::smart_shift * sizeof(T)) ||
+                    !isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
+                {
+                    TransformDispatcher<false>::call(src, dst, op, mask, stream);
+                    return;
+                }
+
+                const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
+                const dim3 grid(divUp(src.cols, threads.x * ft::smart_shift), divUp(src.rows, threads.y), 1);
+
+                transformSmart<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
+                cudaSafeCall( cudaGetLastError() );
+
+                if (stream == 0)
+                    cudaSafeCall( cudaDeviceSynchronize() );
+            }
+
+            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+            static void call(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, Mask mask, cudaStream_t stream)
+            {
+                typedef TransformFunctorTraits<BinOp> ft;
+
+                CV_StaticAssert(ft::smart_shift != 1, "");
+
+                if (!isAligned(src1.data, ft::smart_shift * sizeof(T1)) || !isAligned(src1.step, ft::smart_shift * sizeof(T1)) ||
+                    !isAligned(src2.data, ft::smart_shift * sizeof(T2)) || !isAligned(src2.step, ft::smart_shift * sizeof(T2)) ||
+                    !isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
+                {
+                    TransformDispatcher<false>::call(src1, src2, dst, op, mask, stream);
+                    return;
+                }
+
+                const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
+                const dim3 grid(divUp(src1.cols, threads.x * ft::smart_shift), divUp(src1.rows, threads.y), 1);
+
+                transformSmart<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
+                cudaSafeCall( cudaGetLastError() );
+
+                if (stream == 0)
+                    cudaSafeCall( cudaDeviceSynchronize() );
+            }
+        };
+    } // namespace transform_detail
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_TRANSFORM_DETAIL_HPP__

3rdparty/include/opencv2/core/cuda/detail/type_traits_detail.hpp

@@ -0,0 +1,191 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_TYPE_TRAITS_DETAIL_HPP__
+#define __OPENCV_CUDA_TYPE_TRAITS_DETAIL_HPP__
+
+#include "../common.hpp"
+#include "../vec_traits.hpp"
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    namespace type_traits_detail
+    {
+        template <bool, typename T1, typename T2> struct Select { typedef T1 type; };
+        template <typename T1, typename T2> struct Select<false, T1, T2> { typedef T2 type; };
+
+        template <typename T> struct IsSignedIntergral { enum {value = 0}; };
+        template <> struct IsSignedIntergral<schar> { enum {value = 1}; };
+        template <> struct IsSignedIntergral<char1> { enum {value = 1}; };
+        template <> struct IsSignedIntergral<short> { enum {value = 1}; };
+        template <> struct IsSignedIntergral<short1> { enum {value = 1}; };
+        template <> struct IsSignedIntergral<int> { enum {value = 1}; };
+        template <> struct IsSignedIntergral<int1> { enum {value = 1}; };
+
+        template <typename T> struct IsUnsignedIntegral { enum {value = 0}; };
+        template <> struct IsUnsignedIntegral<uchar> { enum {value = 1}; };
+        template <> struct IsUnsignedIntegral<uchar1> { enum {value = 1}; };
+        template <> struct IsUnsignedIntegral<ushort> { enum {value = 1}; };
+        template <> struct IsUnsignedIntegral<ushort1> { enum {value = 1}; };
+        template <> struct IsUnsignedIntegral<uint> { enum {value = 1}; };
+        template <> struct IsUnsignedIntegral<uint1> { enum {value = 1}; };
+
+        template <typename T> struct IsIntegral { enum {value = IsSignedIntergral<T>::value || IsUnsignedIntegral<T>::value}; };
+        template <> struct IsIntegral<char> { enum {value = 1}; };
+        template <> struct IsIntegral<bool> { enum {value = 1}; };
+
+        template <typename T> struct IsFloat { enum {value = 0}; };
+        template <> struct IsFloat<float> { enum {value = 1}; };
+        template <> struct IsFloat<double> { enum {value = 1}; };
+
+        template <typename T> struct IsVec { enum {value = 0}; };
+        template <> struct IsVec<uchar1> { enum {value = 1}; };
+        template <> struct IsVec<uchar2> { enum {value = 1}; };
+        template <> struct IsVec<uchar3> { enum {value = 1}; };
+        template <> struct IsVec<uchar4> { enum {value = 1}; };
+        template <> struct IsVec<uchar8> { enum {value = 1}; };
+        template <> struct IsVec<char1> { enum {value = 1}; };
+        template <> struct IsVec<char2> { enum {value = 1}; };
+        template <> struct IsVec<char3> { enum {value = 1}; };
+        template <> struct IsVec<char4> { enum {value = 1}; };
+        template <> struct IsVec<char8> { enum {value = 1}; };
+        template <> struct IsVec<ushort1> { enum {value = 1}; };
+        template <> struct IsVec<ushort2> { enum {value = 1}; };
+        template <> struct IsVec<ushort3> { enum {value = 1}; };
+        template <> struct IsVec<ushort4> { enum {value = 1}; };
+        template <> struct IsVec<ushort8> { enum {value = 1}; };
+        template <> struct IsVec<short1> { enum {value = 1}; };
+        template <> struct IsVec<short2> { enum {value = 1}; };
+        template <> struct IsVec<short3> { enum {value = 1}; };
+        template <> struct IsVec<short4> { enum {value = 1}; };
+        template <> struct IsVec<short8> { enum {value = 1}; };
+        template <> struct IsVec<uint1> { enum {value = 1}; };
+        template <> struct IsVec<uint2> { enum {value = 1}; };
+        template <> struct IsVec<uint3> { enum {value = 1}; };
+        template <> struct IsVec<uint4> { enum {value = 1}; };
+        template <> struct IsVec<uint8> { enum {value = 1}; };
+        template <> struct IsVec<int1> { enum {value = 1}; };
+        template <> struct IsVec<int2> { enum {value = 1}; };
+        template <> struct IsVec<int3> { enum {value = 1}; };
+        template <> struct IsVec<int4> { enum {value = 1}; };
+        template <> struct IsVec<int8> { enum {value = 1}; };
+        template <> struct IsVec<float1> { enum {value = 1}; };
+        template <> struct IsVec<float2> { enum {value = 1}; };
+        template <> struct IsVec<float3> { enum {value = 1}; };
+        template <> struct IsVec<float4> { enum {value = 1}; };
+        template <> struct IsVec<float8> { enum {value = 1}; };
+        template <> struct IsVec<double1> { enum {value = 1}; };
+        template <> struct IsVec<double2> { enum {value = 1}; };
+        template <> struct IsVec<double3> { enum {value = 1}; };
+        template <> struct IsVec<double4> { enum {value = 1}; };
+        template <> struct IsVec<double8> { enum {value = 1}; };
+
+        template <class U> struct AddParameterType { typedef const U& type; };
+        template <class U> struct AddParameterType<U&> { typedef U& type; };
+        template <> struct AddParameterType<void> { typedef void type; };
+
+        template <class U> struct ReferenceTraits
+        {
+            enum { value = false };
+            typedef U type;
+        };
+        template <class U> struct ReferenceTraits<U&>
+        {
+            enum { value = true };
+            typedef U type;
+        };
+
+        template <class U> struct PointerTraits
+        {
+            enum { value = false };
+            typedef void type;
+        };
+        template <class U> struct PointerTraits<U*>
+        {
+            enum { value = true };
+            typedef U type;
+        };
+        template <class U> struct PointerTraits<U*&>
+        {
+            enum { value = true };
+            typedef U type;
+        };
+
+        template <class U> struct UnConst
+        {
+            typedef U type;
+            enum { value = 0 };
+        };
+        template <class U> struct UnConst<const U>
+        {
+            typedef U type;
+            enum { value = 1 };
+        };
+        template <class U> struct UnConst<const U&>
+        {
+            typedef U& type;
+            enum { value = 1 };
+        };
+
+        template <class U> struct UnVolatile
+        {
+            typedef U type;
+            enum { value = 0 };
+        };
+        template <class U> struct UnVolatile<volatile U>
+        {
+            typedef U type;
+            enum { value = 1 };
+        };
+        template <class U> struct UnVolatile<volatile U&>
+        {
+            typedef U& type;
+            enum { value = 1 };
+        };
+    } // namespace type_traits_detail
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_TYPE_TRAITS_DETAIL_HPP__

3rdparty/include/opencv2/core/cuda/detail/vec_distance_detail.hpp

@@ -0,0 +1,121 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_VEC_DISTANCE_DETAIL_HPP__
+#define __OPENCV_CUDA_VEC_DISTANCE_DETAIL_HPP__
+
+#include "../datamov_utils.hpp"
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    namespace vec_distance_detail
+    {
+        template <int THREAD_DIM, int N> struct UnrollVecDiffCached
+        {
+            template <typename Dist, typename T1, typename T2>
+            static __device__ void calcCheck(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int ind)
+            {
+                if (ind < len)
+                {
+                    T1 val1 = *vecCached++;
+
+                    T2 val2;
+                    ForceGlob<T2>::Load(vecGlob, ind, val2);
+
+                    dist.reduceIter(val1, val2);
+
+                    UnrollVecDiffCached<THREAD_DIM, N - 1>::calcCheck(vecCached, vecGlob, len, dist, ind + THREAD_DIM);
+                }
+            }
+
+            template <typename Dist, typename T1, typename T2>
+            static __device__ void calcWithoutCheck(const T1* vecCached, const T2* vecGlob, Dist& dist)
+            {
+                T1 val1 = *vecCached++;
+
+                T2 val2;
+                ForceGlob<T2>::Load(vecGlob, 0, val2);
+                vecGlob += THREAD_DIM;
+
+                dist.reduceIter(val1, val2);
+
+                UnrollVecDiffCached<THREAD_DIM, N - 1>::calcWithoutCheck(vecCached, vecGlob, dist);
+            }
+        };
+        template <int THREAD_DIM> struct UnrollVecDiffCached<THREAD_DIM, 0>
+        {
+            template <typename Dist, typename T1, typename T2>
+            static __device__ __forceinline__ void calcCheck(const T1*, const T2*, int, Dist&, int)
+            {
+            }
+
+            template <typename Dist, typename T1, typename T2>
+            static __device__ __forceinline__ void calcWithoutCheck(const T1*, const T2*, Dist&)
+            {
+            }
+        };
+
+        template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN> struct VecDiffCachedCalculator;
+        template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, false>
+        {
+            template <typename Dist, typename T1, typename T2>
+            static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
+            {
+                UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcCheck(vecCached, vecGlob, len, dist, tid);
+            }
+        };
+        template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, true>
+        {
+            template <typename Dist, typename T1, typename T2>
+            static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
+            {
+                UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcWithoutCheck(vecCached, vecGlob + tid, dist);
+            }
+        };
+    } // namespace vec_distance_detail
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_VEC_DISTANCE_DETAIL_HPP__

3rdparty/include/opencv2/core/cuda/dynamic_smem.hpp

@@ -0,0 +1,88 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_DYNAMIC_SMEM_HPP__
+#define __OPENCV_CUDA_DYNAMIC_SMEM_HPP__
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template<class T> struct DynamicSharedMem
+    {
+        __device__ __forceinline__ operator T*()
+        {
+            extern __shared__ int __smem[];
+            return (T*)__smem;
+        }
+
+        __device__ __forceinline__ operator const T*() const
+        {
+            extern __shared__ int __smem[];
+            return (T*)__smem;
+        }
+    };
+
+    // specialize for double to avoid unaligned memory access compile errors
+    template<> struct DynamicSharedMem<double>
+    {
+        __device__ __forceinline__ operator double*()
+        {
+            extern __shared__ double __smem_d[];
+            return (double*)__smem_d;
+        }
+
+        __device__ __forceinline__ operator const double*() const
+        {
+            extern __shared__ double __smem_d[];
+            return (double*)__smem_d;
+        }
+    };
+}}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_DYNAMIC_SMEM_HPP__

3rdparty/include/opencv2/core/cuda/emulation.hpp

@@ -0,0 +1,269 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef OPENCV_CUDA_EMULATION_HPP_
+#define OPENCV_CUDA_EMULATION_HPP_
+
+#include "common.hpp"
+#include "warp_reduce.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    struct Emulation
+    {
+
+        static __device__ __forceinline__ int syncthreadsOr(int pred)
+        {
+#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ < 200)
+                // just campilation stab
+                return 0;
+#else
+                return __syncthreads_or(pred);
+#endif
+        }
+
+        template<int CTA_SIZE>
+        static __forceinline__ __device__ int Ballot(int predicate)
+        {
+#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ >= 200)
+            return __ballot(predicate);
+#else
+            __shared__ volatile int cta_buffer[CTA_SIZE];
+
+            int tid = threadIdx.x;
+            cta_buffer[tid] = predicate ? (1 << (tid & 31)) : 0;
+            return warp_reduce(cta_buffer);
+#endif
+        }
+
+        struct smem
+        {
+            enum { TAG_MASK = (1U << ( (sizeof(unsigned int) << 3) - 5U)) - 1U };
+
+            template<typename T>
+            static __device__ __forceinline__ T atomicInc(T* address, T val)
+            {
+#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ < 120)
+                T count;
+                unsigned int tag = threadIdx.x << ( (sizeof(unsigned int) << 3) - 5U);
+                do
+                {
+                    count = *address & TAG_MASK;
+                    count = tag | (count + 1);
+                    *address = count;
+                } while (*address != count);
+
+                return (count & TAG_MASK) - 1;
+#else
+                return ::atomicInc(address, val);
+#endif
+            }
+
+            template<typename T>
+            static __device__ __forceinline__ T atomicAdd(T* address, T val)
+            {
+#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ < 120)
+                T count;
+                unsigned int tag = threadIdx.x << ( (sizeof(unsigned int) << 3) - 5U);
+                do
+                {
+                    count = *address & TAG_MASK;
+                    count = tag | (count + val);
+                    *address = count;
+                } while (*address != count);
+
+                return (count & TAG_MASK) - val;
+#else
+                return ::atomicAdd(address, val);
+#endif
+            }
+
+            template<typename T>
+            static __device__ __forceinline__ T atomicMin(T* address, T val)
+            {
+#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ < 120)
+                T count = ::min(*address, val);
+                do
+                {
+                    *address = count;
+                } while (*address > count);
+
+                return count;
+#else
+                return ::atomicMin(address, val);
+#endif
+            }
+        }; // struct cmem
+
+        struct glob
+        {
+            static __device__ __forceinline__ int atomicAdd(int* address, int val)
+            {
+                return ::atomicAdd(address, val);
+            }
+            static __device__ __forceinline__ unsigned int atomicAdd(unsigned int* address, unsigned int val)
+            {
+                return ::atomicAdd(address, val);
+            }
+            static __device__ __forceinline__ float atomicAdd(float* address, float val)
+            {
+            #if __CUDA_ARCH__ >= 200
+                return ::atomicAdd(address, val);
+            #else
+                int* address_as_i = (int*) address;
+                int old = *address_as_i, assumed;
+                do {
+                    assumed = old;
+                    old = ::atomicCAS(address_as_i, assumed,
+                        __float_as_int(val + __int_as_float(assumed)));
+                } while (assumed != old);
+                return __int_as_float(old);
+            #endif
+            }
+            static __device__ __forceinline__ double atomicAdd(double* address, double val)
+            {
+            #if __CUDA_ARCH__ >= 130
+                unsigned long long int* address_as_ull = (unsigned long long int*) address;
+                unsigned long long int old = *address_as_ull, assumed;
+                do {
+                    assumed = old;
+                    old = ::atomicCAS(address_as_ull, assumed,
+                        __double_as_longlong(val + __longlong_as_double(assumed)));
+                } while (assumed != old);
+                return __longlong_as_double(old);
+            #else
+                (void) address;
+                (void) val;
+                return 0.0;
+            #endif
+            }
+
+            static __device__ __forceinline__ int atomicMin(int* address, int val)
+            {
+                return ::atomicMin(address, val);
+            }
+            static __device__ __forceinline__ float atomicMin(float* address, float val)
+            {
+            #if __CUDA_ARCH__ >= 120
+                int* address_as_i = (int*) address;
+                int old = *address_as_i, assumed;
+                do {
+                    assumed = old;
+                    old = ::atomicCAS(address_as_i, assumed,
+                        __float_as_int(::fminf(val, __int_as_float(assumed))));
+                } while (assumed != old);
+                return __int_as_float(old);
+            #else
+                (void) address;
+                (void) val;
+                return 0.0f;
+            #endif
+            }
+            static __device__ __forceinline__ double atomicMin(double* address, double val)
+            {
+            #if __CUDA_ARCH__ >= 130
+                unsigned long long int* address_as_ull = (unsigned long long int*) address;
+                unsigned long long int old = *address_as_ull, assumed;
+                do {
+                    assumed = old;
+                    old = ::atomicCAS(address_as_ull, assumed,
+                        __double_as_longlong(::fmin(val, __longlong_as_double(assumed))));
+                } while (assumed != old);
+                return __longlong_as_double(old);
+            #else
+                (void) address;
+                (void) val;
+                return 0.0;
+            #endif
+            }
+
+            static __device__ __forceinline__ int atomicMax(int* address, int val)
+            {
+                return ::atomicMax(address, val);
+            }
+            static __device__ __forceinline__ float atomicMax(float* address, float val)
+            {
+            #if __CUDA_ARCH__ >= 120
+                int* address_as_i = (int*) address;
+                int old = *address_as_i, assumed;
+                do {
+                    assumed = old;
+                    old = ::atomicCAS(address_as_i, assumed,
+                        __float_as_int(::fmaxf(val, __int_as_float(assumed))));
+                } while (assumed != old);
+                return __int_as_float(old);
+            #else
+                (void) address;
+                (void) val;
+                return 0.0f;
+            #endif
+            }
+            static __device__ __forceinline__ double atomicMax(double* address, double val)
+            {
+            #if __CUDA_ARCH__ >= 130
+                unsigned long long int* address_as_ull = (unsigned long long int*) address;
+                unsigned long long int old = *address_as_ull, assumed;
+                do {
+                    assumed = old;
+                    old = ::atomicCAS(address_as_ull, assumed,
+                        __double_as_longlong(::fmax(val, __longlong_as_double(assumed))));
+                } while (assumed != old);
+                return __longlong_as_double(old);
+            #else
+                (void) address;
+                (void) val;
+                return 0.0;
+            #endif
+            }
+        };
+    }; //struct Emulation
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif /* OPENCV_CUDA_EMULATION_HPP_ */

3rdparty/include/opencv2/core/cuda/filters.hpp

@@ -0,0 +1,286 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_FILTERS_HPP__
+#define __OPENCV_CUDA_FILTERS_HPP__
+
+#include "saturate_cast.hpp"
+#include "vec_traits.hpp"
+#include "vec_math.hpp"
+#include "type_traits.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template <typename Ptr2D> struct PointFilter
+    {
+        typedef typename Ptr2D::elem_type elem_type;
+        typedef float index_type;
+
+        explicit __host__ __device__ __forceinline__ PointFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
+        : src(src_)
+        {
+            (void)fx;
+            (void)fy;
+        }
+
+        __device__ __forceinline__ elem_type operator ()(float y, float x) const
+        {
+            return src(__float2int_rz(y), __float2int_rz(x));
+        }
+
+        Ptr2D src;
+    };
+
+    template <typename Ptr2D> struct LinearFilter
+    {
+        typedef typename Ptr2D::elem_type elem_type;
+        typedef float index_type;
+
+        explicit __host__ __device__ __forceinline__ LinearFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
+        : src(src_)
+        {
+            (void)fx;
+            (void)fy;
+        }
+        __device__ __forceinline__ elem_type operator ()(float y, float x) const
+        {
+            typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
+
+            work_type out = VecTraits<work_type>::all(0);
+
+            const int x1 = __float2int_rd(x);
+            const int y1 = __float2int_rd(y);
+            const int x2 = x1 + 1;
+            const int y2 = y1 + 1;
+
+            elem_type src_reg = src(y1, x1);
+            out = out + src_reg * ((x2 - x) * (y2 - y));
+
+            src_reg = src(y1, x2);
+            out = out + src_reg * ((x - x1) * (y2 - y));
+
+            src_reg = src(y2, x1);
+            out = out + src_reg * ((x2 - x) * (y - y1));
+
+            src_reg = src(y2, x2);
+            out = out + src_reg * ((x - x1) * (y - y1));
+
+            return saturate_cast<elem_type>(out);
+        }
+
+        Ptr2D src;
+    };
+
+    template <typename Ptr2D> struct CubicFilter
+    {
+        typedef typename Ptr2D::elem_type elem_type;
+        typedef float index_type;
+        typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
+
+        explicit __host__ __device__ __forceinline__ CubicFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
+        : src(src_)
+        {
+            (void)fx;
+            (void)fy;
+        }
+
+        static __device__ __forceinline__ float bicubicCoeff(float x_)
+        {
+            float x = fabsf(x_);
+            if (x <= 1.0f)
+            {
+                return x * x * (1.5f * x - 2.5f) + 1.0f;
+            }
+            else if (x < 2.0f)
+            {
+                return x * (x * (-0.5f * x + 2.5f) - 4.0f) + 2.0f;
+            }
+            else
+            {
+                return 0.0f;
+            }
+        }
+
+        __device__ elem_type operator ()(float y, float x) const
+        {
+            const float xmin = ::ceilf(x - 2.0f);
+            const float xmax = ::floorf(x + 2.0f);
+
+            const float ymin = ::ceilf(y - 2.0f);
+            const float ymax = ::floorf(y + 2.0f);
+
+            work_type sum = VecTraits<work_type>::all(0);
+            float wsum = 0.0f;
+
+            for (float cy = ymin; cy <= ymax; cy += 1.0f)
+            {
+                for (float cx = xmin; cx <= xmax; cx += 1.0f)
+                {
+                    const float w = bicubicCoeff(x - cx) * bicubicCoeff(y - cy);
+                    sum = sum + w * src(__float2int_rd(cy), __float2int_rd(cx));
+                    wsum += w;
+                }
+            }
+
+            work_type res = (!wsum)? VecTraits<work_type>::all(0) : sum / wsum;
+
+            return saturate_cast<elem_type>(res);
+        }
+
+        Ptr2D src;
+    };
+    // for integer scaling
+    template <typename Ptr2D> struct IntegerAreaFilter
+    {
+        typedef typename Ptr2D::elem_type elem_type;
+        typedef float index_type;
+
+        explicit __host__ __device__ __forceinline__ IntegerAreaFilter(const Ptr2D& src_, float scale_x_, float scale_y_)
+            : src(src_), scale_x(scale_x_), scale_y(scale_y_), scale(1.f / (scale_x * scale_y)) {}
+
+        __device__ __forceinline__ elem_type operator ()(float y, float x) const
+        {
+            float fsx1 = x * scale_x;
+            float fsx2 = fsx1 + scale_x;
+
+            int sx1 = __float2int_ru(fsx1);
+            int sx2 = __float2int_rd(fsx2);
+
+            float fsy1 = y * scale_y;
+            float fsy2 = fsy1 + scale_y;
+
+            int sy1 = __float2int_ru(fsy1);
+            int sy2 = __float2int_rd(fsy2);
+
+            typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
+            work_type out = VecTraits<work_type>::all(0.f);
+
+            for(int dy = sy1; dy < sy2; ++dy)
+                for(int dx = sx1; dx < sx2; ++dx)
+                {
+                    out = out + src(dy, dx) * scale;
+                }
+
+            return saturate_cast<elem_type>(out);
+        }
+
+        Ptr2D src;
+        float scale_x, scale_y ,scale;
+    };
+
+    template <typename Ptr2D> struct AreaFilter
+    {
+        typedef typename Ptr2D::elem_type elem_type;
+        typedef float index_type;
+
+        explicit __host__ __device__ __forceinline__ AreaFilter(const Ptr2D& src_, float scale_x_, float scale_y_)
+            : src(src_), scale_x(scale_x_), scale_y(scale_y_){}
+
+        __device__ __forceinline__ elem_type operator ()(float y, float x) const
+        {
+            float fsx1 = x * scale_x;
+            float fsx2 = fsx1 + scale_x;
+
+            int sx1 = __float2int_ru(fsx1);
+            int sx2 = __float2int_rd(fsx2);
+
+            float fsy1 = y * scale_y;
+            float fsy2 = fsy1 + scale_y;
+
+            int sy1 = __float2int_ru(fsy1);
+            int sy2 = __float2int_rd(fsy2);
+
+            float scale = 1.f / (fminf(scale_x, src.width - fsx1) * fminf(scale_y, src.height - fsy1));
+
+            typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
+            work_type out = VecTraits<work_type>::all(0.f);
+
+            for (int dy = sy1; dy < sy2; ++dy)
+            {
+                for (int dx = sx1; dx < sx2; ++dx)
+                    out = out + src(dy, dx) * scale;
+
+                if (sx1 > fsx1)
+                    out = out + src(dy, (sx1 -1) ) * ((sx1 - fsx1) * scale);
+
+                if (sx2 < fsx2)
+                    out = out + src(dy, sx2) * ((fsx2 -sx2) * scale);
+            }
+
+            if (sy1 > fsy1)
+                for (int dx = sx1; dx < sx2; ++dx)
+                    out = out + src( (sy1 - 1) , dx) * ((sy1 -fsy1) * scale);
+
+            if (sy2 < fsy2)
+                for (int dx = sx1; dx < sx2; ++dx)
+                    out = out + src(sy2, dx) * ((fsy2 -sy2) * scale);
+
+            if ((sy1 > fsy1) &&  (sx1 > fsx1))
+                out = out + src( (sy1 - 1) , (sx1 - 1)) * ((sy1 -fsy1) * (sx1 -fsx1) * scale);
+
+            if ((sy1 > fsy1) &&  (sx2 < fsx2))
+                out = out + src( (sy1 - 1) , sx2) * ((sy1 -fsy1) * (fsx2 -sx2) * scale);
+
+            if ((sy2 < fsy2) &&  (sx2 < fsx2))
+                out = out + src(sy2, sx2) * ((fsy2 -sy2) * (fsx2 -sx2) * scale);
+
+            if ((sy2 < fsy2) &&  (sx1 > fsx1))
+                out = out + src(sy2, (sx1 - 1)) * ((fsy2 -sy2) * (sx1 -fsx1) * scale);
+
+            return saturate_cast<elem_type>(out);
+        }
+
+        Ptr2D src;
+        float scale_x, scale_y;
+        int width, haight;
+    };
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_FILTERS_HPP__

3rdparty/include/opencv2/core/cuda/funcattrib.hpp

@@ -0,0 +1,79 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_DEVICE_FUNCATTRIB_HPP_
+#define __OPENCV_CUDA_DEVICE_FUNCATTRIB_HPP_
+
+#include <cstdio>
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template<class Func>
+    void printFuncAttrib(Func& func)
+    {
+
+        cudaFuncAttributes attrs;
+        cudaFuncGetAttributes(&attrs, func);
+
+        printf("=== Function stats ===\n");
+        printf("Name: \n");
+        printf("sharedSizeBytes    = %d\n", attrs.sharedSizeBytes);
+        printf("constSizeBytes     = %d\n", attrs.constSizeBytes);
+        printf("localSizeBytes     = %d\n", attrs.localSizeBytes);
+        printf("maxThreadsPerBlock = %d\n", attrs.maxThreadsPerBlock);
+        printf("numRegs            = %d\n", attrs.numRegs);
+        printf("ptxVersion         = %d\n", attrs.ptxVersion);
+        printf("binaryVersion      = %d\n", attrs.binaryVersion);
+        printf("\n");
+        fflush(stdout);
+    }
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif  /* __OPENCV_CUDA_DEVICE_FUNCATTRIB_HPP_ */

3rdparty/include/opencv2/core/cuda/functional.hpp

@@ -0,0 +1,797 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_FUNCTIONAL_HPP__
+#define __OPENCV_CUDA_FUNCTIONAL_HPP__
+
+#include <functional>
+#include "saturate_cast.hpp"
+#include "vec_traits.hpp"
+#include "type_traits.hpp"
+#include "device_functions.h"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    // Function Objects
+    template<typename Argument, typename Result> struct unary_function : public std::unary_function<Argument, Result> {};
+    template<typename Argument1, typename Argument2, typename Result> struct binary_function : public std::binary_function<Argument1, Argument2, Result> {};
+
+    // Arithmetic Operations
+    template <typename T> struct plus : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
+                                                 typename TypeTraits<T>::ParameterType b) const
+        {
+            return a + b;
+        }
+        __host__ __device__ __forceinline__ plus() {}
+        __host__ __device__ __forceinline__ plus(const plus&) {}
+    };
+
+    template <typename T> struct minus : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
+                                                 typename TypeTraits<T>::ParameterType b) const
+        {
+            return a - b;
+        }
+        __host__ __device__ __forceinline__ minus() {}
+        __host__ __device__ __forceinline__ minus(const minus&) {}
+    };
+
+    template <typename T> struct multiplies : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
+                                                 typename TypeTraits<T>::ParameterType b) const
+        {
+            return a * b;
+        }
+        __host__ __device__ __forceinline__ multiplies() {}
+        __host__ __device__ __forceinline__ multiplies(const multiplies&) {}
+    };
+
+    template <typename T> struct divides : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
+                                                 typename TypeTraits<T>::ParameterType b) const
+        {
+            return a / b;
+        }
+        __host__ __device__ __forceinline__ divides() {}
+        __host__ __device__ __forceinline__ divides(const divides&) {}
+    };
+
+    template <typename T> struct modulus : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
+                                                 typename TypeTraits<T>::ParameterType b) const
+        {
+            return a % b;
+        }
+        __host__ __device__ __forceinline__ modulus() {}
+        __host__ __device__ __forceinline__ modulus(const modulus&) {}
+    };
+
+    template <typename T> struct negate : unary_function<T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a) const
+        {
+            return -a;
+        }
+        __host__ __device__ __forceinline__ negate() {}
+        __host__ __device__ __forceinline__ negate(const negate&) {}
+    };
+
+    // Comparison Operations
+    template <typename T> struct equal_to : binary_function<T, T, bool>
+    {
+        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
+                                                    typename TypeTraits<T>::ParameterType b) const
+        {
+            return a == b;
+        }
+        __host__ __device__ __forceinline__ equal_to() {}
+        __host__ __device__ __forceinline__ equal_to(const equal_to&) {}
+    };
+
+    template <typename T> struct not_equal_to : binary_function<T, T, bool>
+    {
+        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
+                                                    typename TypeTraits<T>::ParameterType b) const
+        {
+            return a != b;
+        }
+        __host__ __device__ __forceinline__ not_equal_to() {}
+        __host__ __device__ __forceinline__ not_equal_to(const not_equal_to&) {}
+    };
+
+    template <typename T> struct greater : binary_function<T, T, bool>
+    {
+        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
+                                                    typename TypeTraits<T>::ParameterType b) const
+        {
+            return a > b;
+        }
+        __host__ __device__ __forceinline__ greater() {}
+        __host__ __device__ __forceinline__ greater(const greater&) {}
+    };
+
+    template <typename T> struct less : binary_function<T, T, bool>
+    {
+        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
+                                                    typename TypeTraits<T>::ParameterType b) const
+        {
+            return a < b;
+        }
+        __host__ __device__ __forceinline__ less() {}
+        __host__ __device__ __forceinline__ less(const less&) {}
+    };
+
+    template <typename T> struct greater_equal : binary_function<T, T, bool>
+    {
+        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
+                                                    typename TypeTraits<T>::ParameterType b) const
+        {
+            return a >= b;
+        }
+        __host__ __device__ __forceinline__ greater_equal() {}
+        __host__ __device__ __forceinline__ greater_equal(const greater_equal&) {}
+    };
+
+    template <typename T> struct less_equal : binary_function<T, T, bool>
+    {
+        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
+                                                    typename TypeTraits<T>::ParameterType b) const
+        {
+            return a <= b;
+        }
+        __host__ __device__ __forceinline__ less_equal() {}
+        __host__ __device__ __forceinline__ less_equal(const less_equal&) {}
+    };
+
+    // Logical Operations
+    template <typename T> struct logical_and : binary_function<T, T, bool>
+    {
+        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
+                                                    typename TypeTraits<T>::ParameterType b) const
+        {
+            return a && b;
+        }
+        __host__ __device__ __forceinline__ logical_and() {}
+        __host__ __device__ __forceinline__ logical_and(const logical_and&) {}
+    };
+
+    template <typename T> struct logical_or : binary_function<T, T, bool>
+    {
+        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
+                                                    typename TypeTraits<T>::ParameterType b) const
+        {
+            return a || b;
+        }
+        __host__ __device__ __forceinline__ logical_or() {}
+        __host__ __device__ __forceinline__ logical_or(const logical_or&) {}
+    };
+
+    template <typename T> struct logical_not : unary_function<T, bool>
+    {
+        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a) const
+        {
+            return !a;
+        }
+        __host__ __device__ __forceinline__ logical_not() {}
+        __host__ __device__ __forceinline__ logical_not(const logical_not&) {}
+    };
+
+    // Bitwise Operations
+    template <typename T> struct bit_and : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
+                                                 typename TypeTraits<T>::ParameterType b) const
+        {
+            return a & b;
+        }
+        __host__ __device__ __forceinline__ bit_and() {}
+        __host__ __device__ __forceinline__ bit_and(const bit_and&) {}
+    };
+
+    template <typename T> struct bit_or : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
+                                                 typename TypeTraits<T>::ParameterType b) const
+        {
+            return a | b;
+        }
+        __host__ __device__ __forceinline__ bit_or() {}
+        __host__ __device__ __forceinline__ bit_or(const bit_or&) {}
+    };
+
+    template <typename T> struct bit_xor : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
+                                                 typename TypeTraits<T>::ParameterType b) const
+        {
+            return a ^ b;
+        }
+        __host__ __device__ __forceinline__ bit_xor() {}
+        __host__ __device__ __forceinline__ bit_xor(const bit_xor&) {}
+    };
+
+    template <typename T> struct bit_not : unary_function<T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType v) const
+        {
+            return ~v;
+        }
+        __host__ __device__ __forceinline__ bit_not() {}
+        __host__ __device__ __forceinline__ bit_not(const bit_not&) {}
+    };
+
+    // Generalized Identity Operations
+    template <typename T> struct identity : unary_function<T, T>
+    {
+        __device__ __forceinline__ typename TypeTraits<T>::ParameterType operator()(typename TypeTraits<T>::ParameterType x) const
+        {
+            return x;
+        }
+        __host__ __device__ __forceinline__ identity() {}
+        __host__ __device__ __forceinline__ identity(const identity&) {}
+    };
+
+    template <typename T1, typename T2> struct project1st : binary_function<T1, T2, T1>
+    {
+        __device__ __forceinline__ typename TypeTraits<T1>::ParameterType operator()(typename TypeTraits<T1>::ParameterType lhs, typename TypeTraits<T2>::ParameterType rhs) const
+        {
+            return lhs;
+        }
+        __host__ __device__ __forceinline__ project1st() {}
+        __host__ __device__ __forceinline__ project1st(const project1st&) {}
+    };
+
+    template <typename T1, typename T2> struct project2nd : binary_function<T1, T2, T2>
+    {
+        __device__ __forceinline__ typename TypeTraits<T2>::ParameterType operator()(typename TypeTraits<T1>::ParameterType lhs, typename TypeTraits<T2>::ParameterType rhs) const
+        {
+            return rhs;
+        }
+        __host__ __device__ __forceinline__ project2nd() {}
+        __host__ __device__ __forceinline__ project2nd(const project2nd&) {}
+    };
+
+    // Min/Max Operations
+
+#define OPENCV_CUDA_IMPLEMENT_MINMAX(name, type, op) \
+    template <> struct name<type> : binary_function<type, type, type> \
+    { \
+        __device__ __forceinline__ type operator()(type lhs, type rhs) const {return op(lhs, rhs);} \
+        __host__ __device__ __forceinline__ name() {}\
+        __host__ __device__ __forceinline__ name(const name&) {}\
+    };
+
+    template <typename T> struct maximum : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType lhs, typename TypeTraits<T>::ParameterType rhs) const
+        {
+            return max(lhs, rhs);
+        }
+        __host__ __device__ __forceinline__ maximum() {}
+        __host__ __device__ __forceinline__ maximum(const maximum&) {}
+    };
+
+    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, uchar, ::max)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, schar, ::max)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, char, ::max)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, ushort, ::max)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, short, ::max)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, int, ::max)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, uint, ::max)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, float, ::fmax)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, double, ::fmax)
+
+    template <typename T> struct minimum : binary_function<T, T, T>
+    {
+        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType lhs, typename TypeTraits<T>::ParameterType rhs) const
+        {
+            return min(lhs, rhs);
+        }
+        __host__ __device__ __forceinline__ minimum() {}
+        __host__ __device__ __forceinline__ minimum(const minimum&) {}
+    };
+
+    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, uchar, ::min)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, schar, ::min)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, char, ::min)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, ushort, ::min)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, short, ::min)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, int, ::min)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, uint, ::min)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, float, ::fmin)
+    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, double, ::fmin)
+
+#undef OPENCV_CUDA_IMPLEMENT_MINMAX
+
+    // Math functions
+
+    template <typename T> struct abs_func : unary_function<T, T>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType x) const
+        {
+            return abs(x);
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+    template <> struct abs_func<unsigned char> : unary_function<unsigned char, unsigned char>
+    {
+        __device__ __forceinline__ unsigned char operator ()(unsigned char x) const
+        {
+            return x;
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+    template <> struct abs_func<signed char> : unary_function<signed char, signed char>
+    {
+        __device__ __forceinline__ signed char operator ()(signed char x) const
+        {
+            return ::abs((int)x);
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+    template <> struct abs_func<char> : unary_function<char, char>
+    {
+        __device__ __forceinline__ char operator ()(char x) const
+        {
+            return ::abs((int)x);
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+    template <> struct abs_func<unsigned short> : unary_function<unsigned short, unsigned short>
+    {
+        __device__ __forceinline__ unsigned short operator ()(unsigned short x) const
+        {
+            return x;
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+    template <> struct abs_func<short> : unary_function<short, short>
+    {
+        __device__ __forceinline__ short operator ()(short x) const
+        {
+            return ::abs((int)x);
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+    template <> struct abs_func<unsigned int> : unary_function<unsigned int, unsigned int>
+    {
+        __device__ __forceinline__ unsigned int operator ()(unsigned int x) const
+        {
+            return x;
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+    template <> struct abs_func<int> : unary_function<int, int>
+    {
+        __device__ __forceinline__ int operator ()(int x) const
+        {
+            return ::abs(x);
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+    template <> struct abs_func<float> : unary_function<float, float>
+    {
+        __device__ __forceinline__ float operator ()(float x) const
+        {
+            return ::fabsf(x);
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+    template <> struct abs_func<double> : unary_function<double, double>
+    {
+        __device__ __forceinline__ double operator ()(double x) const
+        {
+            return ::fabs(x);
+        }
+
+        __host__ __device__ __forceinline__ abs_func() {}
+        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
+    };
+
+#define OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(name, func) \
+    template <typename T> struct name ## _func : unary_function<T, float> \
+    { \
+        __device__ __forceinline__ float operator ()(typename TypeTraits<T>::ParameterType v) const \
+        { \
+            return func ## f(v); \
+        } \
+        __host__ __device__ __forceinline__ name ## _func() {} \
+        __host__ __device__ __forceinline__ name ## _func(const name ## _func&) {} \
+    }; \
+    template <> struct name ## _func<double> : unary_function<double, double> \
+    { \
+        __device__ __forceinline__ double operator ()(double v) const \
+        { \
+            return func(v); \
+        } \
+        __host__ __device__ __forceinline__ name ## _func() {} \
+        __host__ __device__ __forceinline__ name ## _func(const name ## _func&) {} \
+    };
+
+#define OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR(name, func) \
+    template <typename T> struct name ## _func : binary_function<T, T, float> \
+    { \
+        __device__ __forceinline__ float operator ()(typename TypeTraits<T>::ParameterType v1, typename TypeTraits<T>::ParameterType v2) const \
+        { \
+            return func ## f(v1, v2); \
+        } \
+        __host__ __device__ __forceinline__ name ## _func() {} \
+        __host__ __device__ __forceinline__ name ## _func(const name ## _func&) {} \
+    }; \
+    template <> struct name ## _func<double> : binary_function<double, double, double> \
+    { \
+        __device__ __forceinline__ double operator ()(double v1, double v2) const \
+        { \
+            return func(v1, v2); \
+        } \
+        __host__ __device__ __forceinline__ name ## _func() {} \
+        __host__ __device__ __forceinline__ name ## _func(const name ## _func&) {} \
+    };
+
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(sqrt, ::sqrt)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(exp, ::exp)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(exp2, ::exp2)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(exp10, ::exp10)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(log, ::log)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(log2, ::log2)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(log10, ::log10)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(sin, ::sin)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(cos, ::cos)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(tan, ::tan)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(asin, ::asin)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(acos, ::acos)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(atan, ::atan)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(sinh, ::sinh)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(cosh, ::cosh)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(tanh, ::tanh)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(asinh, ::asinh)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(acosh, ::acosh)
+    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(atanh, ::atanh)
+
+    OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR(hypot, ::hypot)
+    OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR(atan2, ::atan2)
+    OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR(pow, ::pow)
+
+    #undef OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR
+    #undef OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR_NO_DOUBLE
+    #undef OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR
+
+    template<typename T> struct hypot_sqr_func : binary_function<T, T, float>
+    {
+        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType src1, typename TypeTraits<T>::ParameterType src2) const
+        {
+            return src1 * src1 + src2 * src2;
+        }
+        __host__ __device__ __forceinline__ hypot_sqr_func() {}
+        __host__ __device__ __forceinline__ hypot_sqr_func(const hypot_sqr_func&) {}
+    };
+
+    // Saturate Cast Functor
+    template <typename T, typename D> struct saturate_cast_func : unary_function<T, D>
+    {
+        __device__ __forceinline__ D operator ()(typename TypeTraits<T>::ParameterType v) const
+        {
+            return saturate_cast<D>(v);
+        }
+        __host__ __device__ __forceinline__ saturate_cast_func() {}
+        __host__ __device__ __forceinline__ saturate_cast_func(const saturate_cast_func&) {}
+    };
+
+    // Threshold Functors
+    template <typename T> struct thresh_binary_func : unary_function<T, T>
+    {
+        __host__ __device__ __forceinline__ thresh_binary_func(T thresh_, T maxVal_) : thresh(thresh_), maxVal(maxVal_) {}
+
+        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
+        {
+            return (src > thresh) * maxVal;
+        }
+
+        __host__ __device__ __forceinline__ thresh_binary_func() {}
+        __host__ __device__ __forceinline__ thresh_binary_func(const thresh_binary_func& other)
+            : thresh(other.thresh), maxVal(other.maxVal) {}
+
+        T thresh;
+        T maxVal;
+    };
+
+    template <typename T> struct thresh_binary_inv_func : unary_function<T, T>
+    {
+        __host__ __device__ __forceinline__ thresh_binary_inv_func(T thresh_, T maxVal_) : thresh(thresh_), maxVal(maxVal_) {}
+
+        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
+        {
+            return (src <= thresh) * maxVal;
+        }
+
+        __host__ __device__ __forceinline__ thresh_binary_inv_func() {}
+        __host__ __device__ __forceinline__ thresh_binary_inv_func(const thresh_binary_inv_func& other)
+            : thresh(other.thresh), maxVal(other.maxVal) {}
+
+        T thresh;
+        T maxVal;
+    };
+
+    template <typename T> struct thresh_trunc_func : unary_function<T, T>
+    {
+        explicit __host__ __device__ __forceinline__ thresh_trunc_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {(void)maxVal_;}
+
+        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
+        {
+            return minimum<T>()(src, thresh);
+        }
+
+        __host__ __device__ __forceinline__ thresh_trunc_func() {}
+        __host__ __device__ __forceinline__ thresh_trunc_func(const thresh_trunc_func& other)
+            : thresh(other.thresh) {}
+
+        T thresh;
+    };
+
+    template <typename T> struct thresh_to_zero_func : unary_function<T, T>
+    {
+        explicit __host__ __device__ __forceinline__ thresh_to_zero_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {(void)maxVal_;}
+
+        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
+        {
+            return (src > thresh) * src;
+        }
+
+        __host__ __device__ __forceinline__ thresh_to_zero_func() {}
+       __host__  __device__ __forceinline__ thresh_to_zero_func(const thresh_to_zero_func& other)
+            : thresh(other.thresh) {}
+
+        T thresh;
+    };
+
+    template <typename T> struct thresh_to_zero_inv_func : unary_function<T, T>
+    {
+        explicit __host__ __device__ __forceinline__ thresh_to_zero_inv_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {(void)maxVal_;}
+
+        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
+        {
+            return (src <= thresh) * src;
+        }
+
+        __host__ __device__ __forceinline__ thresh_to_zero_inv_func() {}
+        __host__ __device__ __forceinline__ thresh_to_zero_inv_func(const thresh_to_zero_inv_func& other)
+            : thresh(other.thresh) {}
+
+        T thresh;
+    };
+
+    // Function Object Adaptors
+    template <typename Predicate> struct unary_negate : unary_function<typename Predicate::argument_type, bool>
+    {
+      explicit __host__ __device__ __forceinline__ unary_negate(const Predicate& p) : pred(p) {}
+
+      __device__ __forceinline__ bool operator()(typename TypeTraits<typename Predicate::argument_type>::ParameterType x) const
+      {
+          return !pred(x);
+      }
+
+      __host__ __device__ __forceinline__ unary_negate() {}
+      __host__ __device__ __forceinline__ unary_negate(const unary_negate& other) : pred(other.pred) {}
+
+      Predicate pred;
+    };
+
+    template <typename Predicate> __host__ __device__ __forceinline__ unary_negate<Predicate> not1(const Predicate& pred)
+    {
+        return unary_negate<Predicate>(pred);
+    }
+
+    template <typename Predicate> struct binary_negate : binary_function<typename Predicate::first_argument_type, typename Predicate::second_argument_type, bool>
+    {
+        explicit __host__ __device__ __forceinline__ binary_negate(const Predicate& p) : pred(p) {}
+
+        __device__ __forceinline__ bool operator()(typename TypeTraits<typename Predicate::first_argument_type>::ParameterType x,
+                                                   typename TypeTraits<typename Predicate::second_argument_type>::ParameterType y) const
+        {
+            return !pred(x,y);
+        }
+
+        __host__ __device__ __forceinline__ binary_negate() {}
+        __host__ __device__ __forceinline__ binary_negate(const binary_negate& other) : pred(other.pred) {}
+
+        Predicate pred;
+    };
+
+    template <typename BinaryPredicate> __host__ __device__ __forceinline__ binary_negate<BinaryPredicate> not2(const BinaryPredicate& pred)
+    {
+        return binary_negate<BinaryPredicate>(pred);
+    }
+
+    template <typename Op> struct binder1st : unary_function<typename Op::second_argument_type, typename Op::result_type>
+    {
+        __host__ __device__ __forceinline__ binder1st(const Op& op_, const typename Op::first_argument_type& arg1_) : op(op_), arg1(arg1_) {}
+
+        __device__ __forceinline__ typename Op::result_type operator ()(typename TypeTraits<typename Op::second_argument_type>::ParameterType a) const
+        {
+            return op(arg1, a);
+        }
+
+        __host__ __device__ __forceinline__ binder1st() {}
+        __host__ __device__ __forceinline__ binder1st(const binder1st& other) : op(other.op), arg1(other.arg1) {}
+
+        Op op;
+        typename Op::first_argument_type arg1;
+    };
+
+    template <typename Op, typename T> __host__ __device__ __forceinline__ binder1st<Op> bind1st(const Op& op, const T& x)
+    {
+        return binder1st<Op>(op, typename Op::first_argument_type(x));
+    }
+
+    template <typename Op> struct binder2nd : unary_function<typename Op::first_argument_type, typename Op::result_type>
+    {
+        __host__ __device__ __forceinline__ binder2nd(const Op& op_, const typename Op::second_argument_type& arg2_) : op(op_), arg2(arg2_) {}
+
+        __forceinline__ __device__ typename Op::result_type operator ()(typename TypeTraits<typename Op::first_argument_type>::ParameterType a) const
+        {
+            return op(a, arg2);
+        }
+
+        __host__ __device__ __forceinline__ binder2nd() {}
+        __host__ __device__ __forceinline__ binder2nd(const binder2nd& other) : op(other.op), arg2(other.arg2) {}
+
+        Op op;
+        typename Op::second_argument_type arg2;
+    };
+
+    template <typename Op, typename T> __host__ __device__ __forceinline__ binder2nd<Op> bind2nd(const Op& op, const T& x)
+    {
+        return binder2nd<Op>(op, typename Op::second_argument_type(x));
+    }
+
+    // Functor Traits
+    template <typename F> struct IsUnaryFunction
+    {
+        typedef char Yes;
+        struct No {Yes a[2];};
+
+        template <typename T, typename D> static Yes check(unary_function<T, D>);
+        static No check(...);
+
+        static F makeF();
+
+        enum { value = (sizeof(check(makeF())) == sizeof(Yes)) };
+    };
+
+    template <typename F> struct IsBinaryFunction
+    {
+        typedef char Yes;
+        struct No {Yes a[2];};
+
+        template <typename T1, typename T2, typename D> static Yes check(binary_function<T1, T2, D>);
+        static No check(...);
+
+        static F makeF();
+
+        enum { value = (sizeof(check(makeF())) == sizeof(Yes)) };
+    };
+
+    namespace functional_detail
+    {
+        template <size_t src_elem_size, size_t dst_elem_size> struct UnOpShift { enum { shift = 1 }; };
+        template <size_t src_elem_size> struct UnOpShift<src_elem_size, 1> { enum { shift = 4 }; };
+        template <size_t src_elem_size> struct UnOpShift<src_elem_size, 2> { enum { shift = 2 }; };
+
+        template <typename T, typename D> struct DefaultUnaryShift
+        {
+            enum { shift = UnOpShift<sizeof(T), sizeof(D)>::shift };
+        };
+
+        template <size_t src_elem_size1, size_t src_elem_size2, size_t dst_elem_size> struct BinOpShift { enum { shift = 1 }; };
+        template <size_t src_elem_size1, size_t src_elem_size2> struct BinOpShift<src_elem_size1, src_elem_size2, 1> { enum { shift = 4 }; };
+        template <size_t src_elem_size1, size_t src_elem_size2> struct BinOpShift<src_elem_size1, src_elem_size2, 2> { enum { shift = 2 }; };
+
+        template <typename T1, typename T2, typename D> struct DefaultBinaryShift
+        {
+            enum { shift = BinOpShift<sizeof(T1), sizeof(T2), sizeof(D)>::shift };
+        };
+
+        template <typename Func, bool unary = IsUnaryFunction<Func>::value> struct ShiftDispatcher;
+        template <typename Func> struct ShiftDispatcher<Func, true>
+        {
+            enum { shift = DefaultUnaryShift<typename Func::argument_type, typename Func::result_type>::shift };
+        };
+        template <typename Func> struct ShiftDispatcher<Func, false>
+        {
+            enum { shift = DefaultBinaryShift<typename Func::first_argument_type, typename Func::second_argument_type, typename Func::result_type>::shift };
+        };
+    }
+
+    template <typename Func> struct DefaultTransformShift
+    {
+        enum { shift = functional_detail::ShiftDispatcher<Func>::shift };
+    };
+
+    template <typename Func> struct DefaultTransformFunctorTraits
+    {
+        enum { simple_block_dim_x = 16 };
+        enum { simple_block_dim_y = 16 };
+
+        enum { smart_block_dim_x = 16 };
+        enum { smart_block_dim_y = 16 };
+        enum { smart_shift = DefaultTransformShift<Func>::shift };
+    };
+
+    template <typename Func> struct TransformFunctorTraits : DefaultTransformFunctorTraits<Func> {};
+
+#define OPENCV_CUDA_TRANSFORM_FUNCTOR_TRAITS(type) \
+    template <> struct TransformFunctorTraits< type > : DefaultTransformFunctorTraits< type >
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_FUNCTIONAL_HPP__

3rdparty/include/opencv2/core/cuda/limits.hpp

@@ -0,0 +1,128 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_LIMITS_HPP__
+#define __OPENCV_CUDA_LIMITS_HPP__
+
+#include <limits.h>
+#include <float.h>
+#include "common.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+template <class T> struct numeric_limits;
+
+template <> struct numeric_limits<bool>
+{
+    __device__ __forceinline__ static bool min() { return false; }
+    __device__ __forceinline__ static bool max() { return true;  }
+    static const bool is_signed = false;
+};
+
+template <> struct numeric_limits<signed char>
+{
+    __device__ __forceinline__ static signed char min() { return SCHAR_MIN; }
+    __device__ __forceinline__ static signed char max() { return SCHAR_MAX; }
+    static const bool is_signed = true;
+};
+
+template <> struct numeric_limits<unsigned char>
+{
+    __device__ __forceinline__ static unsigned char min() { return 0; }
+    __device__ __forceinline__ static unsigned char max() { return UCHAR_MAX; }
+    static const bool is_signed = false;
+};
+
+template <> struct numeric_limits<short>
+{
+    __device__ __forceinline__ static short min() { return SHRT_MIN; }
+    __device__ __forceinline__ static short max() { return SHRT_MAX; }
+    static const bool is_signed = true;
+};
+
+template <> struct numeric_limits<unsigned short>
+{
+    __device__ __forceinline__ static unsigned short min() { return 0; }
+    __device__ __forceinline__ static unsigned short max() { return USHRT_MAX; }
+    static const bool is_signed = false;
+};
+
+template <> struct numeric_limits<int>
+{
+    __device__ __forceinline__ static int min() { return INT_MIN; }
+    __device__ __forceinline__ static int max() { return INT_MAX; }
+    static const bool is_signed = true;
+};
+
+template <> struct numeric_limits<unsigned int>
+{
+    __device__ __forceinline__ static unsigned int min() { return 0; }
+    __device__ __forceinline__ static unsigned int max() { return UINT_MAX; }
+    static const bool is_signed = false;
+};
+
+template <> struct numeric_limits<float>
+{
+    __device__ __forceinline__ static float min() { return FLT_MIN; }
+    __device__ __forceinline__ static float max() { return FLT_MAX; }
+    __device__ __forceinline__ static float epsilon() { return FLT_EPSILON; }
+    static const bool is_signed = true;
+};
+
+template <> struct numeric_limits<double>
+{
+    __device__ __forceinline__ static double min() { return DBL_MIN; }
+    __device__ __forceinline__ static double max() { return DBL_MAX; }
+    __device__ __forceinline__ static double epsilon() { return DBL_EPSILON; }
+    static const bool is_signed = true;
+};
+}}} // namespace cv { namespace cuda { namespace cudev {
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_LIMITS_HPP__

3rdparty/include/opencv2/core/cuda/reduce.hpp

@@ -0,0 +1,205 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_REDUCE_HPP__
+#define __OPENCV_CUDA_REDUCE_HPP__
+
+#include <thrust/tuple.h>
+#include "detail/reduce.hpp"
+#include "detail/reduce_key_val.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template <int N, typename T, class Op>
+    __device__ __forceinline__ void reduce(volatile T* smem, T& val, unsigned int tid, const Op& op)
+    {
+        reduce_detail::Dispatcher<N>::reductor::template reduce<volatile T*, T&, const Op&>(smem, val, tid, op);
+    }
+    template <int N,
+              typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
+              typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9,
+              class Op0, class Op1, class Op2, class Op3, class Op4, class Op5, class Op6, class Op7, class Op8, class Op9>
+    __device__ __forceinline__ void reduce(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
+                                           const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
+                                           unsigned int tid,
+                                           const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>& op)
+    {
+        reduce_detail::Dispatcher<N>::reductor::template reduce<
+                const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>&,
+                const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>&,
+                const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>&>(smem, val, tid, op);
+    }
+
+    template <unsigned int N, typename K, typename V, class Cmp>
+    __device__ __forceinline__ void reduceKeyVal(volatile K* skeys, K& key, volatile V* svals, V& val, unsigned int tid, const Cmp& cmp)
+    {
+        reduce_key_val_detail::Dispatcher<N>::reductor::template reduce<volatile K*, K&, volatile V*, V&, const Cmp&>(skeys, key, svals, val, tid, cmp);
+    }
+    template <unsigned int N,
+              typename K,
+              typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
+              typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
+              class Cmp>
+    __device__ __forceinline__ void reduceKeyVal(volatile K* skeys, K& key,
+                                                 const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
+                                                 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
+                                                 unsigned int tid, const Cmp& cmp)
+    {
+        reduce_key_val_detail::Dispatcher<N>::reductor::template reduce<volatile K*, K&,
+                const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>&,
+                const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>&,
+                const Cmp&>(skeys, key, svals, val, tid, cmp);
+    }
+    template <unsigned int N,
+              typename KP0, typename KP1, typename KP2, typename KP3, typename KP4, typename KP5, typename KP6, typename KP7, typename KP8, typename KP9,
+              typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9,
+              typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
+              typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
+              class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9>
+    __device__ __forceinline__ void reduceKeyVal(const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>& skeys,
+                                                 const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key,
+                                                 const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
+                                                 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
+                                                 unsigned int tid,
+                                                 const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp)
+    {
+        reduce_key_val_detail::Dispatcher<N>::reductor::template reduce<
+                const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>&,
+                const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>&,
+                const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>&,
+                const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>&,
+                const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>&
+                >(skeys, key, svals, val, tid, cmp);
+    }
+
+    // smem_tuple
+
+    template <typename T0>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*>
+    smem_tuple(T0* t0)
+    {
+        return thrust::make_tuple((volatile T0*) t0);
+    }
+
+    template <typename T0, typename T1>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*, volatile T1*>
+    smem_tuple(T0* t0, T1* t1)
+    {
+        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1);
+    }
+
+    template <typename T0, typename T1, typename T2>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*, volatile T1*, volatile T2*>
+    smem_tuple(T0* t0, T1* t1, T2* t2)
+    {
+        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2);
+    }
+
+    template <typename T0, typename T1, typename T2, typename T3>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*>
+    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3)
+    {
+        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3);
+    }
+
+    template <typename T0, typename T1, typename T2, typename T3, typename T4>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*>
+    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4)
+    {
+        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4);
+    }
+
+    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*>
+    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5)
+    {
+        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5);
+    }
+
+    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*>
+    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6)
+    {
+        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6);
+    }
+
+    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*, volatile T7*>
+    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6, T7* t7)
+    {
+        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6, (volatile T7*) t7);
+    }
+
+    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7, typename T8>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*, volatile T7*, volatile T8*>
+    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6, T7* t7, T8* t8)
+    {
+        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6, (volatile T7*) t7, (volatile T8*) t8);
+    }
+
+    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7, typename T8, typename T9>
+    __device__ __forceinline__
+    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*, volatile T7*, volatile T8*, volatile T9*>
+    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6, T7* t7, T8* t8, T9* t9)
+    {
+        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6, (volatile T7*) t7, (volatile T8*) t8, (volatile T9*) t9);
+    }
+}}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_UTILITY_HPP__

3rdparty/include/opencv2/core/cuda/saturate_cast.hpp

@@ -0,0 +1,292 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_SATURATE_CAST_HPP__
+#define __OPENCV_CUDA_SATURATE_CAST_HPP__
+
+#include "common.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uchar v) { return _Tp(v); }
+    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(schar v) { return _Tp(v); }
+    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(ushort v) { return _Tp(v); }
+    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(short v) { return _Tp(v); }
+    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uint v) { return _Tp(v); }
+    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(int v) { return _Tp(v); }
+    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(float v) { return _Tp(v); }
+    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(double v) { return _Tp(v); }
+
+    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(schar v)
+    {
+        uint res = 0;
+        int vi = v;
+        asm("cvt.sat.u8.s8 %0, %1;" : "=r"(res) : "r"(vi));
+        return res;
+    }
+    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(short v)
+    {
+        uint res = 0;
+        asm("cvt.sat.u8.s16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(ushort v)
+    {
+        uint res = 0;
+        asm("cvt.sat.u8.u16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(int v)
+    {
+        uint res = 0;
+        asm("cvt.sat.u8.s32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(uint v)
+    {
+        uint res = 0;
+        asm("cvt.sat.u8.u32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(float v)
+    {
+        uint res = 0;
+        asm("cvt.rni.sat.u8.f32 %0, %1;" : "=r"(res) : "f"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(double v)
+    {
+    #if __CUDA_ARCH__ >= 130
+        uint res = 0;
+        asm("cvt.rni.sat.u8.f64 %0, %1;" : "=r"(res) : "d"(v));
+        return res;
+    #else
+        return saturate_cast<uchar>((float)v);
+    #endif
+    }
+
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(uchar v)
+    {
+        uint res = 0;
+        uint vi = v;
+        asm("cvt.sat.s8.u8 %0, %1;" : "=r"(res) : "r"(vi));
+        return res;
+    }
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(short v)
+    {
+        uint res = 0;
+        asm("cvt.sat.s8.s16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(ushort v)
+    {
+        uint res = 0;
+        asm("cvt.sat.s8.u16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(int v)
+    {
+        uint res = 0;
+        asm("cvt.sat.s8.s32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(uint v)
+    {
+        uint res = 0;
+        asm("cvt.sat.s8.u32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(float v)
+    {
+        uint res = 0;
+        asm("cvt.rni.sat.s8.f32 %0, %1;" : "=r"(res) : "f"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(double v)
+    {
+    #if __CUDA_ARCH__ >= 130
+        uint res = 0;
+        asm("cvt.rni.sat.s8.f64 %0, %1;" : "=r"(res) : "d"(v));
+        return res;
+    #else
+        return saturate_cast<schar>((float)v);
+    #endif
+    }
+
+    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(schar v)
+    {
+        ushort res = 0;
+        int vi = v;
+        asm("cvt.sat.u16.s8 %0, %1;" : "=h"(res) : "r"(vi));
+        return res;
+    }
+    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(short v)
+    {
+        ushort res = 0;
+        asm("cvt.sat.u16.s16 %0, %1;" : "=h"(res) : "h"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(int v)
+    {
+        ushort res = 0;
+        asm("cvt.sat.u16.s32 %0, %1;" : "=h"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(uint v)
+    {
+        ushort res = 0;
+        asm("cvt.sat.u16.u32 %0, %1;" : "=h"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(float v)
+    {
+        ushort res = 0;
+        asm("cvt.rni.sat.u16.f32 %0, %1;" : "=h"(res) : "f"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(double v)
+    {
+    #if __CUDA_ARCH__ >= 130
+        ushort res = 0;
+        asm("cvt.rni.sat.u16.f64 %0, %1;" : "=h"(res) : "d"(v));
+        return res;
+    #else
+        return saturate_cast<ushort>((float)v);
+    #endif
+    }
+
+    template<> __device__ __forceinline__ short saturate_cast<short>(ushort v)
+    {
+        short res = 0;
+        asm("cvt.sat.s16.u16 %0, %1;" : "=h"(res) : "h"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ short saturate_cast<short>(int v)
+    {
+        short res = 0;
+        asm("cvt.sat.s16.s32 %0, %1;" : "=h"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ short saturate_cast<short>(uint v)
+    {
+        short res = 0;
+        asm("cvt.sat.s16.u32 %0, %1;" : "=h"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ short saturate_cast<short>(float v)
+    {
+        short res = 0;
+        asm("cvt.rni.sat.s16.f32 %0, %1;" : "=h"(res) : "f"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ short saturate_cast<short>(double v)
+    {
+    #if __CUDA_ARCH__ >= 130
+        short res = 0;
+        asm("cvt.rni.sat.s16.f64 %0, %1;" : "=h"(res) : "d"(v));
+        return res;
+    #else
+        return saturate_cast<short>((float)v);
+    #endif
+    }
+
+    template<> __device__ __forceinline__ int saturate_cast<int>(uint v)
+    {
+        int res = 0;
+        asm("cvt.sat.s32.u32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ int saturate_cast<int>(float v)
+    {
+        return __float2int_rn(v);
+    }
+    template<> __device__ __forceinline__ int saturate_cast<int>(double v)
+    {
+    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
+        return __double2int_rn(v);
+    #else
+        return saturate_cast<int>((float)v);
+    #endif
+    }
+
+    template<> __device__ __forceinline__ uint saturate_cast<uint>(schar v)
+    {
+        uint res = 0;
+        int vi = v;
+        asm("cvt.sat.u32.s8 %0, %1;" : "=r"(res) : "r"(vi));
+        return res;
+    }
+    template<> __device__ __forceinline__ uint saturate_cast<uint>(short v)
+    {
+        uint res = 0;
+        asm("cvt.sat.u32.s16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ uint saturate_cast<uint>(int v)
+    {
+        uint res = 0;
+        asm("cvt.sat.u32.s32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ uint saturate_cast<uint>(float v)
+    {
+        return __float2uint_rn(v);
+    }
+    template<> __device__ __forceinline__ uint saturate_cast<uint>(double v)
+    {
+    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
+        return __double2uint_rn(v);
+    #else
+        return saturate_cast<uint>((float)v);
+    #endif
+    }
+}}}
+
+//! @endcond
+
+#endif /* __OPENCV_CUDA_SATURATE_CAST_HPP__ */

3rdparty/include/opencv2/core/cuda/scan.hpp

@@ -0,0 +1,258 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_SCAN_HPP__
+#define __OPENCV_CUDA_SCAN_HPP__
+
+#include "opencv2/core/cuda/common.hpp"
+#include "opencv2/core/cuda/utility.hpp"
+#include "opencv2/core/cuda/warp.hpp"
+#include "opencv2/core/cuda/warp_shuffle.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    enum ScanKind { EXCLUSIVE = 0,  INCLUSIVE = 1 };
+
+    template <ScanKind Kind, typename T, typename F> struct WarpScan
+    {
+        __device__ __forceinline__ WarpScan() {}
+        __device__ __forceinline__ WarpScan(const WarpScan& other) { (void)other; }
+
+        __device__ __forceinline__ T operator()( volatile T *ptr , const unsigned int idx)
+        {
+            const unsigned int lane = idx & 31;
+            F op;
+
+            if ( lane >=  1) ptr [idx ] = op(ptr [idx -  1], ptr [idx]);
+            if ( lane >=  2) ptr [idx ] = op(ptr [idx -  2], ptr [idx]);
+            if ( lane >=  4) ptr [idx ] = op(ptr [idx -  4], ptr [idx]);
+            if ( lane >=  8) ptr [idx ] = op(ptr [idx -  8], ptr [idx]);
+            if ( lane >= 16) ptr [idx ] = op(ptr [idx - 16], ptr [idx]);
+
+            if( Kind == INCLUSIVE )
+                return ptr [idx];
+            else
+                return (lane > 0) ? ptr [idx - 1] : 0;
+        }
+
+        __device__ __forceinline__ unsigned int index(const unsigned int tid)
+        {
+            return tid;
+        }
+
+        __device__ __forceinline__ void init(volatile T *ptr){}
+
+        static const int warp_offset      = 0;
+
+        typedef WarpScan<INCLUSIVE, T, F>  merge;
+    };
+
+    template <ScanKind Kind , typename T, typename F> struct WarpScanNoComp
+    {
+        __device__ __forceinline__ WarpScanNoComp() {}
+        __device__ __forceinline__ WarpScanNoComp(const WarpScanNoComp& other) { (void)other; }
+
+        __device__ __forceinline__ T operator()( volatile T *ptr , const unsigned int idx)
+        {
+            const unsigned int lane = threadIdx.x & 31;
+            F op;
+
+            ptr [idx ] = op(ptr [idx -  1], ptr [idx]);
+            ptr [idx ] = op(ptr [idx -  2], ptr [idx]);
+            ptr [idx ] = op(ptr [idx -  4], ptr [idx]);
+            ptr [idx ] = op(ptr [idx -  8], ptr [idx]);
+            ptr [idx ] = op(ptr [idx - 16], ptr [idx]);
+
+            if( Kind == INCLUSIVE )
+                return ptr [idx];
+            else
+                return (lane > 0) ? ptr [idx - 1] : 0;
+        }
+
+        __device__ __forceinline__ unsigned int index(const unsigned int tid)
+        {
+            return (tid >> warp_log) * warp_smem_stride + 16 + (tid & warp_mask);
+        }
+
+        __device__ __forceinline__ void init(volatile T *ptr)
+        {
+            ptr[threadIdx.x] = 0;
+        }
+
+        static const int warp_smem_stride = 32 + 16 + 1;
+        static const int warp_offset      = 16;
+        static const int warp_log         = 5;
+        static const int warp_mask        = 31;
+
+        typedef WarpScanNoComp<INCLUSIVE, T, F> merge;
+    };
+
+    template <ScanKind Kind , typename T, typename Sc, typename F> struct BlockScan
+    {
+        __device__ __forceinline__ BlockScan() {}
+        __device__ __forceinline__ BlockScan(const BlockScan& other) { (void)other; }
+
+        __device__ __forceinline__ T operator()(volatile T *ptr)
+        {
+            const unsigned int tid  = threadIdx.x;
+            const unsigned int lane = tid & warp_mask;
+            const unsigned int warp = tid >> warp_log;
+
+            Sc scan;
+            typename Sc::merge merge_scan;
+            const unsigned int idx = scan.index(tid);
+
+            T val = scan(ptr, idx);
+            __syncthreads ();
+
+            if( warp == 0)
+                scan.init(ptr);
+            __syncthreads ();
+
+            if( lane == 31 )
+                ptr [scan.warp_offset + warp ] = (Kind == INCLUSIVE) ? val : ptr [idx];
+            __syncthreads ();
+
+            if( warp == 0 )
+                merge_scan(ptr, idx);
+            __syncthreads();
+
+            if ( warp > 0)
+                val = ptr [scan.warp_offset + warp - 1] + val;
+            __syncthreads ();
+
+            ptr[idx] = val;
+            __syncthreads ();
+
+            return val ;
+        }
+
+        static const int warp_log  = 5;
+        static const int warp_mask = 31;
+    };
+
+    template <typename T>
+    __device__ T warpScanInclusive(T idata, volatile T* s_Data, unsigned int tid)
+    {
+    #if __CUDA_ARCH__ >= 300
+        const unsigned int laneId = cv::cuda::device::Warp::laneId();
+
+        // scan on shuffl functions
+        #pragma unroll
+        for (int i = 1; i <= (OPENCV_CUDA_WARP_SIZE / 2); i *= 2)
+        {
+            const T n = cv::cuda::device::shfl_up(idata, i);
+            if (laneId >= i)
+                  idata += n;
+        }
+
+        return idata;
+    #else
+        unsigned int pos = 2 * tid - (tid & (OPENCV_CUDA_WARP_SIZE - 1));
+        s_Data[pos] = 0;
+        pos += OPENCV_CUDA_WARP_SIZE;
+        s_Data[pos] = idata;
+
+        s_Data[pos] += s_Data[pos - 1];
+        s_Data[pos] += s_Data[pos - 2];
+        s_Data[pos] += s_Data[pos - 4];
+        s_Data[pos] += s_Data[pos - 8];
+        s_Data[pos] += s_Data[pos - 16];
+
+        return s_Data[pos];
+    #endif
+    }
+
+    template <typename T>
+    __device__ __forceinline__ T warpScanExclusive(T idata, volatile T* s_Data, unsigned int tid)
+    {
+        return warpScanInclusive(idata, s_Data, tid) - idata;
+    }
+
+    template <int tiNumScanThreads, typename T>
+    __device__ T blockScanInclusive(T idata, volatile T* s_Data, unsigned int tid)
+    {
+        if (tiNumScanThreads > OPENCV_CUDA_WARP_SIZE)
+        {
+            //Bottom-level inclusive warp scan
+            T warpResult = warpScanInclusive(idata, s_Data, tid);
+
+            //Save top elements of each warp for exclusive warp scan
+            //sync to wait for warp scans to complete (because s_Data is being overwritten)
+            __syncthreads();
+            if ((tid & (OPENCV_CUDA_WARP_SIZE - 1)) == (OPENCV_CUDA_WARP_SIZE - 1))
+            {
+                s_Data[tid >> OPENCV_CUDA_LOG_WARP_SIZE] = warpResult;
+            }
+
+            //wait for warp scans to complete
+            __syncthreads();
+
+            if (tid < (tiNumScanThreads / OPENCV_CUDA_WARP_SIZE) )
+            {
+                //grab top warp elements
+                T val = s_Data[tid];
+                //calculate exclusive scan and write back to shared memory
+                s_Data[tid] = warpScanExclusive(val, s_Data, tid);
+            }
+
+            //return updated warp scans with exclusive scan results
+            __syncthreads();
+
+            return warpResult + s_Data[tid >> OPENCV_CUDA_LOG_WARP_SIZE];
+        }
+        else
+        {
+            return warpScanInclusive(idata, s_Data, tid);
+        }
+    }
+}}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_SCAN_HPP__

3rdparty/include/opencv2/core/cuda/simd_functions.hpp

@@ -0,0 +1,869 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+/*
+ * Copyright (c) 2013 NVIDIA Corporation. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ *   Redistributions of source code must retain the above copyright notice,
+ *   this list of conditions and the following disclaimer.
+ *
+ *   Redistributions in binary form must reproduce the above copyright notice,
+ *   this list of conditions and the following disclaimer in the documentation
+ *   and/or other materials provided with the distribution.
+ *
+ *   Neither the name of NVIDIA Corporation nor the names of its contributors
+ *   may be used to endorse or promote products derived from this software
+ *   without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+ * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+ * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+ * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+ * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+ * POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef __OPENCV_CUDA_SIMD_FUNCTIONS_HPP__
+#define __OPENCV_CUDA_SIMD_FUNCTIONS_HPP__
+
+#include "common.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    // 2
+
+    static __device__ __forceinline__ unsigned int vadd2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vadd2.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vadd.u32.u32.u32.sat %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vadd.u32.u32.u32.sat %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s;
+        s = a ^ b;          // sum bits
+        r = a + b;          // actual sum
+        s = s ^ r;          // determine carry-ins for each bit position
+        s = s & 0x00010000; // carry-in to high word (= carry-out from low word)
+        r = r - s;          // subtract out carry-out from low word
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsub2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vsub2.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vsub.u32.u32.u32.sat %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vsub.u32.u32.u32.sat %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s;
+        s = a ^ b;          // sum bits
+        r = a - b;          // actual sum
+        s = s ^ r;          // determine carry-ins for each bit position
+        s = s & 0x00010000; // borrow to high word
+        r = r + s;          // compensate for borrow from low word
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vabsdiff2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vabsdiff2.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vabsdiff.u32.u32.u32.sat %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vabsdiff.u32.u32.u32.sat %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s, t, u, v;
+        s = a & 0x0000ffff; // extract low halfword
+        r = b & 0x0000ffff; // extract low halfword
+        u = ::max(r, s);    // maximum of low halfwords
+        v = ::min(r, s);    // minimum of low halfwords
+        s = a & 0xffff0000; // extract high halfword
+        r = b & 0xffff0000; // extract high halfword
+        t = ::max(r, s);    // maximum of high halfwords
+        s = ::min(r, s);    // minimum of high halfwords
+        r = u | t;          // maximum of both halfwords
+        s = v | s;          // minimum of both halfwords
+        r = r - s;          // |a - b| = max(a,b) - min(a,b);
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vavg2(unsigned int a, unsigned int b)
+    {
+        unsigned int r, s;
+
+        // HAKMEM #23: a + b = 2 * (a & b) + (a ^ b) ==>
+        // (a + b) / 2 = (a & b) + ((a ^ b) >> 1)
+        s = a ^ b;
+        r = a & b;
+        s = s & 0xfffefffe; // ensure shift doesn't cross halfword boundaries
+        s = s >> 1;
+        s = r + s;
+
+        return s;
+    }
+
+    static __device__ __forceinline__ unsigned int vavrg2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vavrg2.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        // HAKMEM #23: a + b = 2 * (a | b) - (a ^ b) ==>
+        // (a + b + 1) / 2 = (a | b) - ((a ^ b) >> 1)
+        unsigned int s;
+        s = a ^ b;
+        r = a | b;
+        s = s & 0xfffefffe; // ensure shift doesn't cross half-word boundaries
+        s = s >> 1;
+        r = r - s;
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vseteq2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset2.u32.u32.eq %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        // inspired by Alan Mycroft's null-byte detection algorithm:
+        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
+        unsigned int c;
+        r = a ^ b;          // 0x0000 if a == b
+        c = r | 0x80008000; // set msbs, to catch carry out
+        r = r ^ c;          // extract msbs, msb = 1 if r < 0x8000
+        c = c - 0x00010001; // msb = 0, if r was 0x0000 or 0x8000
+        c = r & ~c;         // msb = 1, if r was 0x0000
+        r = c >> 15;        // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmpeq2(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vseteq2(a, b);
+        c = r << 16;        // convert bool
+        r = c - r;          //  into mask
+    #else
+        // inspired by Alan Mycroft's null-byte detection algorithm:
+        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
+        r = a ^ b;          // 0x0000 if a == b
+        c = r | 0x80008000; // set msbs, to catch carry out
+        r = r ^ c;          // extract msbs, msb = 1 if r < 0x8000
+        c = c - 0x00010001; // msb = 0, if r was 0x0000 or 0x8000
+        c = r & ~c;         // msb = 1, if r was 0x0000
+        r = c >> 15;        // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetge2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset2.u32.u32.ge %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int c;
+        asm("not.b32 %0, %0;" : "+r"(b));
+        c = vavrg2(a, b);   // (a + ~b + 1) / 2 = (a - b) / 2
+        c = c & 0x80008000; // msb = carry-outs
+        r = c >> 15;        // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmpge2(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetge2(a, b);
+        c = r << 16;        // convert bool
+        r = c - r;          //  into mask
+    #else
+        asm("not.b32 %0, %0;" : "+r"(b));
+        c = vavrg2(a, b);   // (a + ~b + 1) / 2 = (a - b) / 2
+        c = c & 0x80008000; // msb = carry-outs
+        r = c >> 15;        // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetgt2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset2.u32.u32.gt %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int c;
+        asm("not.b32 %0, %0;" : "+r"(b));
+        c = vavg2(a, b);    // (a + ~b) / 2 = (a - b) / 2 [rounded down]
+        c = c & 0x80008000; // msbs = carry-outs
+        r = c >> 15;        // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmpgt2(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetgt2(a, b);
+        c = r << 16;        // convert bool
+        r = c - r;          //  into mask
+    #else
+        asm("not.b32 %0, %0;" : "+r"(b));
+        c = vavg2(a, b);    // (a + ~b) / 2 = (a - b) / 2 [rounded down]
+        c = c & 0x80008000; // msbs = carry-outs
+        r = c >> 15;        // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetle2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset2.u32.u32.le %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int c;
+        asm("not.b32 %0, %0;" : "+r"(a));
+        c = vavrg2(a, b);   // (b + ~a + 1) / 2 = (b - a) / 2
+        c = c & 0x80008000; // msb = carry-outs
+        r = c >> 15;        // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmple2(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetle2(a, b);
+        c = r << 16;        // convert bool
+        r = c - r;          //  into mask
+    #else
+        asm("not.b32 %0, %0;" : "+r"(a));
+        c = vavrg2(a, b);   // (b + ~a + 1) / 2 = (b - a) / 2
+        c = c & 0x80008000; // msb = carry-outs
+        r = c >> 15;        // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetlt2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset2.u32.u32.lt %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int c;
+        asm("not.b32 %0, %0;" : "+r"(a));
+        c = vavg2(a, b);    // (b + ~a) / 2 = (b - a) / 2 [rounded down]
+        c = c & 0x80008000; // msb = carry-outs
+        r = c >> 15;        // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmplt2(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetlt2(a, b);
+        c = r << 16;        // convert bool
+        r = c - r;          //  into mask
+    #else
+        asm("not.b32 %0, %0;" : "+r"(a));
+        c = vavg2(a, b);    // (b + ~a) / 2 = (b - a) / 2 [rounded down]
+        c = c & 0x80008000; // msb = carry-outs
+        r = c >> 15;        // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetne2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm ("vset2.u32.u32.ne %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        // inspired by Alan Mycroft's null-byte detection algorithm:
+        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
+        unsigned int c;
+        r = a ^ b;          // 0x0000 if a == b
+        c = r | 0x80008000; // set msbs, to catch carry out
+        c = c - 0x00010001; // msb = 0, if r was 0x0000 or 0x8000
+        c = r | c;          // msb = 1, if r was not 0x0000
+        c = c & 0x80008000; // extract msbs
+        r = c >> 15;        // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmpne2(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetne2(a, b);
+        c = r << 16;        // convert bool
+        r = c - r;          //  into mask
+    #else
+        // inspired by Alan Mycroft's null-byte detection algorithm:
+        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
+        r = a ^ b;          // 0x0000 if a == b
+        c = r | 0x80008000; // set msbs, to catch carry out
+        c = c - 0x00010001; // msb = 0, if r was 0x0000 or 0x8000
+        c = r | c;          // msb = 1, if r was not 0x0000
+        c = c & 0x80008000; // extract msbs
+        r = c >> 15;        // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vmax2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vmax2.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vmax.u32.u32.u32 %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vmax.u32.u32.u32 %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s, t, u;
+        r = a & 0x0000ffff; // extract low halfword
+        s = b & 0x0000ffff; // extract low halfword
+        t = ::max(r, s);    // maximum of low halfwords
+        r = a & 0xffff0000; // extract high halfword
+        s = b & 0xffff0000; // extract high halfword
+        u = ::max(r, s);    // maximum of high halfwords
+        r = t | u;          // combine halfword maximums
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vmin2(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vmin2.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vmin.u32.u32.u32 %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vmin.u32.u32.u32 %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s, t, u;
+        r = a & 0x0000ffff; // extract low halfword
+        s = b & 0x0000ffff; // extract low halfword
+        t = ::min(r, s);    // minimum of low halfwords
+        r = a & 0xffff0000; // extract high halfword
+        s = b & 0xffff0000; // extract high halfword
+        u = ::min(r, s);    // minimum of high halfwords
+        r = t | u;          // combine halfword minimums
+    #endif
+
+        return r;
+    }
+
+    // 4
+
+    static __device__ __forceinline__ unsigned int vadd4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vadd4.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vadd.u32.u32.u32.sat %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vadd.u32.u32.u32.sat %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vadd.u32.u32.u32.sat %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vadd.u32.u32.u32.sat %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s, t;
+        s = a ^ b;          // sum bits
+        r = a & 0x7f7f7f7f; // clear msbs
+        t = b & 0x7f7f7f7f; // clear msbs
+        s = s & 0x80808080; // msb sum bits
+        r = r + t;          // add without msbs, record carry-out in msbs
+        r = r ^ s;          // sum of msb sum and carry-in bits, w/o carry-out
+    #endif /* __CUDA_ARCH__ >= 300 */
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsub4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vsub4.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vsub.u32.u32.u32.sat %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vsub.u32.u32.u32.sat %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vsub.u32.u32.u32.sat %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vsub.u32.u32.u32.sat %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s, t;
+        s = a ^ ~b;         // inverted sum bits
+        r = a | 0x80808080; // set msbs
+        t = b & 0x7f7f7f7f; // clear msbs
+        s = s & 0x80808080; // inverted msb sum bits
+        r = r - t;          // subtract w/o msbs, record inverted borrows in msb
+        r = r ^ s;          // combine inverted msb sum bits and borrows
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vavg4(unsigned int a, unsigned int b)
+    {
+        unsigned int r, s;
+
+        // HAKMEM #23: a + b = 2 * (a & b) + (a ^ b) ==>
+        // (a + b) / 2 = (a & b) + ((a ^ b) >> 1)
+        s = a ^ b;
+        r = a & b;
+        s = s & 0xfefefefe; // ensure following shift doesn't cross byte boundaries
+        s = s >> 1;
+        s = r + s;
+
+        return s;
+    }
+
+    static __device__ __forceinline__ unsigned int vavrg4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vavrg4.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        // HAKMEM #23: a + b = 2 * (a | b) - (a ^ b) ==>
+        // (a + b + 1) / 2 = (a | b) - ((a ^ b) >> 1)
+        unsigned int c;
+        c = a ^ b;
+        r = a | b;
+        c = c & 0xfefefefe; // ensure following shift doesn't cross byte boundaries
+        c = c >> 1;
+        r = r - c;
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vseteq4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset4.u32.u32.eq %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        // inspired by Alan Mycroft's null-byte detection algorithm:
+        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
+        unsigned int c;
+        r = a ^ b;          // 0x00 if a == b
+        c = r | 0x80808080; // set msbs, to catch carry out
+        r = r ^ c;          // extract msbs, msb = 1 if r < 0x80
+        c = c - 0x01010101; // msb = 0, if r was 0x00 or 0x80
+        c = r & ~c;         // msb = 1, if r was 0x00
+        r = c >> 7;         // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmpeq4(unsigned int a, unsigned int b)
+    {
+        unsigned int r, t;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vseteq4(a, b);
+        t = r << 8;         // convert bool
+        r = t - r;          //  to mask
+    #else
+        // inspired by Alan Mycroft's null-byte detection algorithm:
+        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
+        t = a ^ b;          // 0x00 if a == b
+        r = t | 0x80808080; // set msbs, to catch carry out
+        t = t ^ r;          // extract msbs, msb = 1 if t < 0x80
+        r = r - 0x01010101; // msb = 0, if t was 0x00 or 0x80
+        r = t & ~r;         // msb = 1, if t was 0x00
+        t = r >> 7;         // build mask
+        t = r - t;          //  from
+        r = t | r;          //   msbs
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetle4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset4.u32.u32.le %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int c;
+        asm("not.b32 %0, %0;" : "+r"(a));
+        c = vavrg4(a, b);   // (b + ~a + 1) / 2 = (b - a) / 2
+        c = c & 0x80808080; // msb = carry-outs
+        r = c >> 7;         // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmple4(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetle4(a, b);
+        c = r << 8;         // convert bool
+        r = c - r;          //  to mask
+    #else
+        asm("not.b32 %0, %0;" : "+r"(a));
+        c = vavrg4(a, b);   // (b + ~a + 1) / 2 = (b - a) / 2
+        c = c & 0x80808080; // msbs = carry-outs
+        r = c >> 7;         // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetlt4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset4.u32.u32.lt %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int c;
+        asm("not.b32 %0, %0;" : "+r"(a));
+        c = vavg4(a, b);    // (b + ~a) / 2 = (b - a) / 2 [rounded down]
+        c = c & 0x80808080; // msb = carry-outs
+        r = c >> 7;         // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmplt4(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetlt4(a, b);
+        c = r << 8;         // convert bool
+        r = c - r;          //  to mask
+    #else
+        asm("not.b32 %0, %0;" : "+r"(a));
+        c = vavg4(a, b);    // (b + ~a) / 2 = (b - a) / 2 [rounded down]
+        c = c & 0x80808080; // msbs = carry-outs
+        r = c >> 7;         // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetge4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset4.u32.u32.ge %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int c;
+        asm("not.b32 %0, %0;" : "+r"(b));
+        c = vavrg4(a, b);   // (a + ~b + 1) / 2 = (a - b) / 2
+        c = c & 0x80808080; // msb = carry-outs
+        r = c >> 7;         // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmpge4(unsigned int a, unsigned int b)
+    {
+        unsigned int r, s;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetge4(a, b);
+        s = r << 8;         // convert bool
+        r = s - r;          //  to mask
+    #else
+        asm ("not.b32 %0,%0;" : "+r"(b));
+        r = vavrg4 (a, b);  // (a + ~b + 1) / 2 = (a - b) / 2
+        r = r & 0x80808080; // msb = carry-outs
+        s = r >> 7;         // build mask
+        s = r - s;          //  from
+        r = s | r;          //   msbs
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetgt4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset4.u32.u32.gt %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int c;
+        asm("not.b32 %0, %0;" : "+r"(b));
+        c = vavg4(a, b);    // (a + ~b) / 2 = (a - b) / 2 [rounded down]
+        c = c & 0x80808080; // msb = carry-outs
+        r = c >> 7;         // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmpgt4(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetgt4(a, b);
+        c = r << 8;         // convert bool
+        r = c - r;          //  to mask
+    #else
+        asm("not.b32 %0, %0;" : "+r"(b));
+        c = vavg4(a, b);    // (a + ~b) / 2 = (a - b) / 2 [rounded down]
+        c = c & 0x80808080; // msb = carry-outs
+        r = c >> 7;         // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vsetne4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vset4.u32.u32.ne %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        // inspired by Alan Mycroft's null-byte detection algorithm:
+        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
+        unsigned int c;
+        r = a ^ b;          // 0x00 if a == b
+        c = r | 0x80808080; // set msbs, to catch carry out
+        c = c - 0x01010101; // msb = 0, if r was 0x00 or 0x80
+        c = r | c;          // msb = 1, if r was not 0x00
+        c = c & 0x80808080; // extract msbs
+        r = c >> 7;         // convert to bool
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vcmpne4(unsigned int a, unsigned int b)
+    {
+        unsigned int r, c;
+
+    #if __CUDA_ARCH__ >= 300
+        r = vsetne4(a, b);
+        c = r << 8;         // convert bool
+        r = c - r;          //  to mask
+    #else
+        // inspired by Alan Mycroft's null-byte detection algorithm:
+        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
+        r = a ^ b;          // 0x00 if a == b
+        c = r | 0x80808080; // set msbs, to catch carry out
+        c = c - 0x01010101; // msb = 0, if r was 0x00 or 0x80
+        c = r | c;          // msb = 1, if r was not 0x00
+        c = c & 0x80808080; // extract msbs
+        r = c >> 7;         // convert
+        r = c - r;          //  msbs to
+        r = c | r;          //   mask
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vabsdiff4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vabsdiff4.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vabsdiff.u32.u32.u32.sat %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vabsdiff.u32.u32.u32.sat %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vabsdiff.u32.u32.u32.sat %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vabsdiff.u32.u32.u32.sat %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s;
+        s = vcmpge4(a, b);  // mask = 0xff if a >= b
+        r = a ^ b;          //
+        s = (r &  s) ^ b;   // select a when a >= b, else select b => max(a,b)
+        r = s ^ r;          // select a when b >= a, else select b => min(a,b)
+        r = s - r;          // |a - b| = max(a,b) - min(a,b);
+    #endif
+
+        return r;
+    }
+
+    static __device__ __forceinline__ unsigned int vmax4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vmax4.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vmax.u32.u32.u32 %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vmax.u32.u32.u32 %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vmax.u32.u32.u32 %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vmax.u32.u32.u32 %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s;
+        s = vcmpge4(a, b);  // mask = 0xff if a >= b
+        r = a & s;          // select a when b >= a
+        s = b & ~s;         // select b when b < a
+        r = r | s;          // combine byte selections
+    #endif
+
+        return r;           // byte-wise unsigned maximum
+    }
+
+    static __device__ __forceinline__ unsigned int vmin4(unsigned int a, unsigned int b)
+    {
+        unsigned int r = 0;
+
+    #if __CUDA_ARCH__ >= 300
+        asm("vmin4.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #elif __CUDA_ARCH__ >= 200
+        asm("vmin.u32.u32.u32 %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vmin.u32.u32.u32 %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vmin.u32.u32.u32 %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+        asm("vmin.u32.u32.u32 %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
+    #else
+        unsigned int s;
+        s = vcmpge4(b, a);  // mask = 0xff if a >= b
+        r = a & s;          // select a when b >= a
+        s = b & ~s;         // select b when b < a
+        r = r | s;          // combine byte selections
+    #endif
+
+        return r;
+    }
+}}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_SIMD_FUNCTIONS_HPP__

3rdparty/include/opencv2/core/cuda/transform.hpp

@@ -0,0 +1,75 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_TRANSFORM_HPP__
+#define __OPENCV_CUDA_TRANSFORM_HPP__
+
+#include "common.hpp"
+#include "utility.hpp"
+#include "detail/transform_detail.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template <typename T, typename D, typename UnOp, typename Mask>
+    static inline void transform(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, const Mask& mask, cudaStream_t stream)
+    {
+        typedef TransformFunctorTraits<UnOp> ft;
+        transform_detail::TransformDispatcher<VecTraits<T>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src, dst, op, mask, stream);
+    }
+
+    template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
+    static inline void transform(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, const Mask& mask, cudaStream_t stream)
+    {
+        typedef TransformFunctorTraits<BinOp> ft;
+        transform_detail::TransformDispatcher<VecTraits<T1>::cn == 1 && VecTraits<T2>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src1, src2, dst, op, mask, stream);
+    }
+}}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_TRANSFORM_HPP__

3rdparty/include/opencv2/core/cuda/type_traits.hpp

@@ -0,0 +1,90 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_TYPE_TRAITS_HPP__
+#define __OPENCV_CUDA_TYPE_TRAITS_HPP__
+
+#include "detail/type_traits_detail.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template <typename T> struct IsSimpleParameter
+    {
+        enum {value = type_traits_detail::IsIntegral<T>::value || type_traits_detail::IsFloat<T>::value ||
+            type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<T>::type>::value};
+    };
+
+    template <typename T> struct TypeTraits
+    {
+        typedef typename type_traits_detail::UnConst<T>::type                                                NonConstType;
+        typedef typename type_traits_detail::UnVolatile<T>::type                                             NonVolatileType;
+        typedef typename type_traits_detail::UnVolatile<typename type_traits_detail::UnConst<T>::type>::type UnqualifiedType;
+        typedef typename type_traits_detail::PointerTraits<UnqualifiedType>::type                            PointeeType;
+        typedef typename type_traits_detail::ReferenceTraits<T>::type                                        ReferredType;
+
+        enum { isConst          = type_traits_detail::UnConst<T>::value };
+        enum { isVolatile       = type_traits_detail::UnVolatile<T>::value };
+
+        enum { isReference      = type_traits_detail::ReferenceTraits<UnqualifiedType>::value };
+        enum { isPointer        = type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<UnqualifiedType>::type>::value };
+
+        enum { isUnsignedInt    = type_traits_detail::IsUnsignedIntegral<UnqualifiedType>::value };
+        enum { isSignedInt      = type_traits_detail::IsSignedIntergral<UnqualifiedType>::value };
+        enum { isIntegral       = type_traits_detail::IsIntegral<UnqualifiedType>::value };
+        enum { isFloat          = type_traits_detail::IsFloat<UnqualifiedType>::value };
+        enum { isArith          = isIntegral || isFloat };
+        enum { isVec            = type_traits_detail::IsVec<UnqualifiedType>::value };
+
+        typedef typename type_traits_detail::Select<IsSimpleParameter<UnqualifiedType>::value,
+            T, typename type_traits_detail::AddParameterType<T>::type>::type ParameterType;
+    };
+}}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_TYPE_TRAITS_HPP__

3rdparty/include/opencv2/core/cuda/utility.hpp

@@ -0,0 +1,221 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_UTILITY_HPP__
+#define __OPENCV_CUDA_UTILITY_HPP__
+
+#include "saturate_cast.hpp"
+#include "datamov_utils.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    #define OPENCV_CUDA_LOG_WARP_SIZE        (5)
+    #define OPENCV_CUDA_WARP_SIZE            (1 << OPENCV_CUDA_LOG_WARP_SIZE)
+    #define OPENCV_CUDA_LOG_MEM_BANKS        ((__CUDA_ARCH__ >= 200) ? 5 : 4) // 32 banks on fermi, 16 on tesla
+    #define OPENCV_CUDA_MEM_BANKS            (1 << OPENCV_CUDA_LOG_MEM_BANKS)
+
+    ///////////////////////////////////////////////////////////////////////////////
+    // swap
+
+    template <typename T> void __device__ __host__ __forceinline__ swap(T& a, T& b)
+    {
+        const T temp = a;
+        a = b;
+        b = temp;
+    }
+
+    ///////////////////////////////////////////////////////////////////////////////
+    // Mask Reader
+
+    struct SingleMask
+    {
+        explicit __host__ __device__ __forceinline__ SingleMask(PtrStepb mask_) : mask(mask_) {}
+        __host__ __device__ __forceinline__ SingleMask(const SingleMask& mask_): mask(mask_.mask){}
+
+        __device__ __forceinline__ bool operator()(int y, int x) const
+        {
+            return mask.ptr(y)[x] != 0;
+        }
+
+        PtrStepb mask;
+    };
+
+    struct SingleMaskChannels
+    {
+        __host__ __device__ __forceinline__ SingleMaskChannels(PtrStepb mask_, int channels_)
+        : mask(mask_), channels(channels_) {}
+        __host__ __device__ __forceinline__ SingleMaskChannels(const SingleMaskChannels& mask_)
+            :mask(mask_.mask), channels(mask_.channels){}
+
+        __device__ __forceinline__ bool operator()(int y, int x) const
+        {
+            return mask.ptr(y)[x / channels] != 0;
+        }
+
+        PtrStepb mask;
+        int channels;
+    };
+
+    struct MaskCollection
+    {
+        explicit __host__ __device__ __forceinline__ MaskCollection(PtrStepb* maskCollection_)
+            : maskCollection(maskCollection_) {}
+
+        __device__ __forceinline__ MaskCollection(const MaskCollection& masks_)
+            : maskCollection(masks_.maskCollection), curMask(masks_.curMask){}
+
+        __device__ __forceinline__ void next()
+        {
+            curMask = *maskCollection++;
+        }
+        __device__ __forceinline__ void setMask(int z)
+        {
+            curMask = maskCollection[z];
+        }
+
+        __device__ __forceinline__ bool operator()(int y, int x) const
+        {
+            uchar val;
+            return curMask.data == 0 || (ForceGlob<uchar>::Load(curMask.ptr(y), x, val), (val != 0));
+        }
+
+        const PtrStepb* maskCollection;
+        PtrStepb curMask;
+    };
+
+    struct WithOutMask
+    {
+        __host__ __device__ __forceinline__ WithOutMask(){}
+        __host__ __device__ __forceinline__ WithOutMask(const WithOutMask&){}
+
+        __device__ __forceinline__ void next() const
+        {
+        }
+        __device__ __forceinline__ void setMask(int) const
+        {
+        }
+
+        __device__ __forceinline__ bool operator()(int, int) const
+        {
+            return true;
+        }
+
+        __device__ __forceinline__ bool operator()(int, int, int) const
+        {
+            return true;
+        }
+
+        static __device__ __forceinline__ bool check(int, int)
+        {
+            return true;
+        }
+
+        static __device__ __forceinline__ bool check(int, int, int)
+        {
+            return true;
+        }
+    };
+
+    ///////////////////////////////////////////////////////////////////////////////
+    // Solve linear system
+
+    // solve 2x2 linear system Ax=b
+    template <typename T> __device__ __forceinline__ bool solve2x2(const T A[2][2], const T b[2], T x[2])
+    {
+        T det = A[0][0] * A[1][1] - A[1][0] * A[0][1];
+
+        if (det != 0)
+        {
+            double invdet = 1.0 / det;
+
+            x[0] = saturate_cast<T>(invdet * (b[0] * A[1][1] - b[1] * A[0][1]));
+
+            x[1] = saturate_cast<T>(invdet * (A[0][0] * b[1] - A[1][0] * b[0]));
+
+            return true;
+        }
+
+        return false;
+    }
+
+    // solve 3x3 linear system Ax=b
+    template <typename T> __device__ __forceinline__ bool solve3x3(const T A[3][3], const T b[3], T x[3])
+    {
+        T det = A[0][0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1])
+              - A[0][1] * (A[1][0] * A[2][2] - A[1][2] * A[2][0])
+              + A[0][2] * (A[1][0] * A[2][1] - A[1][1] * A[2][0]);
+
+        if (det != 0)
+        {
+            double invdet = 1.0 / det;
+
+            x[0] = saturate_cast<T>(invdet *
+                (b[0]    * (A[1][1] * A[2][2] - A[1][2] * A[2][1]) -
+                 A[0][1] * (b[1]    * A[2][2] - A[1][2] * b[2]   ) +
+                 A[0][2] * (b[1]    * A[2][1] - A[1][1] * b[2]   )));
+
+            x[1] = saturate_cast<T>(invdet *
+                (A[0][0] * (b[1]    * A[2][2] - A[1][2] * b[2]   ) -
+                 b[0]    * (A[1][0] * A[2][2] - A[1][2] * A[2][0]) +
+                 A[0][2] * (A[1][0] * b[2]    - b[1]    * A[2][0])));
+
+            x[2] = saturate_cast<T>(invdet *
+                (A[0][0] * (A[1][1] * b[2]    - b[1]    * A[2][1]) -
+                 A[0][1] * (A[1][0] * b[2]    - b[1]    * A[2][0]) +
+                 b[0]    * (A[1][0] * A[2][1] - A[1][1] * A[2][0])));
+
+            return true;
+        }
+
+        return false;
+    }
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_UTILITY_HPP__

3rdparty/include/opencv2/core/cuda/vec_distance.hpp

@@ -0,0 +1,232 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_VEC_DISTANCE_HPP__
+#define __OPENCV_CUDA_VEC_DISTANCE_HPP__
+
+#include "reduce.hpp"
+#include "functional.hpp"
+#include "detail/vec_distance_detail.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template <typename T> struct L1Dist
+    {
+        typedef int value_type;
+        typedef int result_type;
+
+        __device__ __forceinline__ L1Dist() : mySum(0) {}
+
+        __device__ __forceinline__ void reduceIter(int val1, int val2)
+        {
+            mySum = __sad(val1, val2, mySum);
+        }
+
+        template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
+        {
+            reduce<THREAD_DIM>(smem, mySum, tid, plus<int>());
+        }
+
+        __device__ __forceinline__ operator int() const
+        {
+            return mySum;
+        }
+
+        int mySum;
+    };
+    template <> struct L1Dist<float>
+    {
+        typedef float value_type;
+        typedef float result_type;
+
+        __device__ __forceinline__ L1Dist() : mySum(0.0f) {}
+
+        __device__ __forceinline__ void reduceIter(float val1, float val2)
+        {
+            mySum += ::fabs(val1 - val2);
+        }
+
+        template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
+        {
+            reduce<THREAD_DIM>(smem, mySum, tid, plus<float>());
+        }
+
+        __device__ __forceinline__ operator float() const
+        {
+            return mySum;
+        }
+
+        float mySum;
+    };
+
+    struct L2Dist
+    {
+        typedef float value_type;
+        typedef float result_type;
+
+        __device__ __forceinline__ L2Dist() : mySum(0.0f) {}
+
+        __device__ __forceinline__ void reduceIter(float val1, float val2)
+        {
+            float reg = val1 - val2;
+            mySum += reg * reg;
+        }
+
+        template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
+        {
+            reduce<THREAD_DIM>(smem, mySum, tid, plus<float>());
+        }
+
+        __device__ __forceinline__ operator float() const
+        {
+            return sqrtf(mySum);
+        }
+
+        float mySum;
+    };
+
+    struct HammingDist
+    {
+        typedef int value_type;
+        typedef int result_type;
+
+        __device__ __forceinline__ HammingDist() : mySum(0) {}
+
+        __device__ __forceinline__ void reduceIter(int val1, int val2)
+        {
+            mySum += __popc(val1 ^ val2);
+        }
+
+        template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
+        {
+            reduce<THREAD_DIM>(smem, mySum, tid, plus<int>());
+        }
+
+        __device__ __forceinline__ operator int() const
+        {
+            return mySum;
+        }
+
+        int mySum;
+    };
+
+    // calc distance between two vectors in global memory
+    template <int THREAD_DIM, typename Dist, typename T1, typename T2>
+    __device__ void calcVecDiffGlobal(const T1* vec1, const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid)
+    {
+        for (int i = tid; i < len; i += THREAD_DIM)
+        {
+            T1 val1;
+            ForceGlob<T1>::Load(vec1, i, val1);
+
+            T2 val2;
+            ForceGlob<T2>::Load(vec2, i, val2);
+
+            dist.reduceIter(val1, val2);
+        }
+
+        dist.reduceAll<THREAD_DIM>(smem, tid);
+    }
+
+    // calc distance between two vectors, first vector is cached in register or shared memory, second vector is in global memory
+    template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename Dist, typename T1, typename T2>
+    __device__ __forceinline__ void calcVecDiffCached(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, typename Dist::result_type* smem, int tid)
+    {
+        vec_distance_detail::VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>::calc(vecCached, vecGlob, len, dist, tid);
+
+        dist.reduceAll<THREAD_DIM>(smem, tid);
+    }
+
+    // calc distance between two vectors in global memory
+    template <int THREAD_DIM, typename T1> struct VecDiffGlobal
+    {
+        explicit __device__ __forceinline__ VecDiffGlobal(const T1* vec1_, int = 0, void* = 0, int = 0, int = 0)
+        {
+            vec1 = vec1_;
+        }
+
+        template <typename T2, typename Dist>
+        __device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
+        {
+            calcVecDiffGlobal<THREAD_DIM>(vec1, vec2, len, dist, smem, tid);
+        }
+
+        const T1* vec1;
+    };
+
+    // calc distance between two vectors, first vector is cached in register memory, second vector is in global memory
+    template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename U> struct VecDiffCachedRegister
+    {
+        template <typename T1> __device__ __forceinline__ VecDiffCachedRegister(const T1* vec1, int len, U* smem, int glob_tid, int tid)
+        {
+            if (glob_tid < len)
+                smem[glob_tid] = vec1[glob_tid];
+            __syncthreads();
+
+            U* vec1ValsPtr = vec1Vals;
+
+            #pragma unroll
+            for (int i = tid; i < MAX_LEN; i += THREAD_DIM)
+                *vec1ValsPtr++ = smem[i];
+
+            __syncthreads();
+        }
+
+        template <typename T2, typename Dist>
+        __device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
+        {
+            calcVecDiffCached<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>(vec1Vals, vec2, len, dist, smem, tid);
+        }
+
+        U vec1Vals[MAX_LEN / THREAD_DIM];
+    };
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_VEC_DISTANCE_HPP__

3rdparty/include/opencv2/core/cuda/vec_math.hpp

@@ -0,0 +1,930 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_VECMATH_HPP__
+#define __OPENCV_CUDA_VECMATH_HPP__
+
+#include "vec_traits.hpp"
+#include "saturate_cast.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+
+// saturate_cast
+
+namespace vec_math_detail
+{
+    template <int cn, typename VecD> struct SatCastHelper;
+    template <typename VecD> struct SatCastHelper<1, VecD>
+    {
+        template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
+        {
+            typedef typename VecTraits<VecD>::elem_type D;
+            return VecTraits<VecD>::make(saturate_cast<D>(v.x));
+        }
+    };
+    template <typename VecD> struct SatCastHelper<2, VecD>
+    {
+        template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
+        {
+            typedef typename VecTraits<VecD>::elem_type D;
+            return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y));
+        }
+    };
+    template <typename VecD> struct SatCastHelper<3, VecD>
+    {
+        template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
+        {
+            typedef typename VecTraits<VecD>::elem_type D;
+            return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z));
+        }
+    };
+    template <typename VecD> struct SatCastHelper<4, VecD>
+    {
+        template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
+        {
+            typedef typename VecTraits<VecD>::elem_type D;
+            return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z), saturate_cast<D>(v.w));
+        }
+    };
+
+    template <typename VecD, typename VecS> static __device__ __forceinline__ VecD saturate_cast_helper(const VecS& v)
+    {
+        return SatCastHelper<VecTraits<VecD>::cn, VecD>::cast(v);
+    }
+}
+
+template<typename T> static __device__ __forceinline__ T saturate_cast(const uchar1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const char1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const ushort1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const short1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const uint1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const int1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const float1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const double1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+
+template<typename T> static __device__ __forceinline__ T saturate_cast(const uchar2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const char2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const ushort2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const short2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const uint2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const int2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const float2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const double2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+
+template<typename T> static __device__ __forceinline__ T saturate_cast(const uchar3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const char3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const ushort3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const short3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const uint3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const int3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const float3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const double3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+
+template<typename T> static __device__ __forceinline__ T saturate_cast(const uchar4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const char4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const ushort4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const short4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const uint4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const int4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const float4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+template<typename T> static __device__ __forceinline__ T saturate_cast(const double4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
+
+// unary operators
+
+#define CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(op, input_type, output_type) \
+    __device__ __forceinline__ output_type ## 1 operator op(const input_type ## 1 & a) \
+    { \
+        return VecTraits<output_type ## 1>::make(op (a.x)); \
+    } \
+    __device__ __forceinline__ output_type ## 2 operator op(const input_type ## 2 & a) \
+    { \
+        return VecTraits<output_type ## 2>::make(op (a.x), op (a.y)); \
+    } \
+    __device__ __forceinline__ output_type ## 3 operator op(const input_type ## 3 & a) \
+    { \
+        return VecTraits<output_type ## 3>::make(op (a.x), op (a.y), op (a.z)); \
+    } \
+    __device__ __forceinline__ output_type ## 4 operator op(const input_type ## 4 & a) \
+    { \
+        return VecTraits<output_type ## 4>::make(op (a.x), op (a.y), op (a.z), op (a.w)); \
+    }
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, char, char)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, short, short)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, int, int)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, char, uchar)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, ushort, uchar)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, short, uchar)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, int, uchar)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, uint, uchar)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, float, uchar)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, double, uchar)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, char, char)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, ushort, ushort)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, short, short)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, int, int)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, uint, uint)
+
+#undef CV_CUDEV_IMPLEMENT_VEC_UNARY_OP
+
+// unary functions
+
+#define CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(func_name, func, input_type, output_type) \
+    __device__ __forceinline__ output_type ## 1 func_name(const input_type ## 1 & a) \
+    { \
+        return VecTraits<output_type ## 1>::make(func (a.x)); \
+    } \
+    __device__ __forceinline__ output_type ## 2 func_name(const input_type ## 2 & a) \
+    { \
+        return VecTraits<output_type ## 2>::make(func (a.x), func (a.y)); \
+    } \
+    __device__ __forceinline__ output_type ## 3 func_name(const input_type ## 3 & a) \
+    { \
+        return VecTraits<output_type ## 3>::make(func (a.x), func (a.y), func (a.z)); \
+    } \
+    __device__ __forceinline__ output_type ## 4 func_name(const input_type ## 4 & a) \
+    { \
+        return VecTraits<output_type ## 4>::make(func (a.x), func (a.y), func (a.z), func (a.w)); \
+    }
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(abs, /*::abs*/, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(abs, ::abs, char, char)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(abs, /*::abs*/, ushort, ushort)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(abs, ::abs, short, short)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(abs, ::abs, int, int)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(abs, /*::abs*/, uint, uint)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(abs, ::fabsf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(abs, ::fabs, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrt, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::exp, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::log, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sin, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cos, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tan, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asin, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acos, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atan, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinh, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::cosh, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanh, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinh, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acosh, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanh, double, double)
+
+#undef CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC
+
+// binary operators (vec & vec)
+
+#define CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(op, input_type, output_type) \
+    __device__ __forceinline__ output_type ## 1 operator op(const input_type ## 1 & a, const input_type ## 1 & b) \
+    { \
+        return VecTraits<output_type ## 1>::make(a.x op b.x); \
+    } \
+    __device__ __forceinline__ output_type ## 2 operator op(const input_type ## 2 & a, const input_type ## 2 & b) \
+    { \
+        return VecTraits<output_type ## 2>::make(a.x op b.x, a.y op b.y); \
+    } \
+    __device__ __forceinline__ output_type ## 3 operator op(const input_type ## 3 & a, const input_type ## 3 & b) \
+    { \
+        return VecTraits<output_type ## 3>::make(a.x op b.x, a.y op b.y, a.z op b.z); \
+    } \
+    __device__ __forceinline__ output_type ## 4 operator op(const input_type ## 4 & a, const input_type ## 4 & b) \
+    { \
+        return VecTraits<output_type ## 4>::make(a.x op b.x, a.y op b.y, a.z op b.z, a.w op b.w); \
+    }
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, uchar, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, char, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, ushort, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, short, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, int, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, uint, uint)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, float, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, uchar, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, char, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, ushort, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, short, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, int, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, uint, uint)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, float, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, uchar, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, char, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, ushort, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, short, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, int, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, uint, uint)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, float, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, uchar, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, char, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, ushort, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, short, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, int, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, uint, uint)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, float, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, char, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, ushort, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, short, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, int, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, uint, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, float, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, double, uchar)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, char, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, ushort, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, short, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, int, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, uint, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, float, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, double, uchar)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, char, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, ushort, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, short, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, int, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, uint, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, float, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, double, uchar)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, char, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, ushort, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, short, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, int, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, uint, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, float, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, double, uchar)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, char, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, ushort, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, short, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, int, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, uint, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, float, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, double, uchar)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, char, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, ushort, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, short, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, int, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, uint, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, float, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, double, uchar)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, char, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, ushort, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, short, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, int, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, uint, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, float, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, double, uchar)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, char, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, ushort, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, short, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, int, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, uint, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, float, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, double, uchar)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, char, char)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, ushort, ushort)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, short, short)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, int, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, uint, uint)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, char, char)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, ushort, ushort)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, short, short)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, int, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, uint, uint)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, char, char)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, ushort, ushort)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, short, short)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, int, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, uint, uint)
+
+#undef CV_CUDEV_IMPLEMENT_VEC_BINARY_OP
+
+// binary operators (vec & scalar)
+
+#define CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(op, input_type, scalar_type, output_type) \
+    __device__ __forceinline__ output_type ## 1 operator op(const input_type ## 1 & a, scalar_type s) \
+    { \
+        return VecTraits<output_type ## 1>::make(a.x op s); \
+    } \
+    __device__ __forceinline__ output_type ## 1 operator op(scalar_type s, const input_type ## 1 & b) \
+    { \
+        return VecTraits<output_type ## 1>::make(s op b.x); \
+    } \
+    __device__ __forceinline__ output_type ## 2 operator op(const input_type ## 2 & a, scalar_type s) \
+    { \
+        return VecTraits<output_type ## 2>::make(a.x op s, a.y op s); \
+    } \
+    __device__ __forceinline__ output_type ## 2 operator op(scalar_type s, const input_type ## 2 & b) \
+    { \
+        return VecTraits<output_type ## 2>::make(s op b.x, s op b.y); \
+    } \
+    __device__ __forceinline__ output_type ## 3 operator op(const input_type ## 3 & a, scalar_type s) \
+    { \
+        return VecTraits<output_type ## 3>::make(a.x op s, a.y op s, a.z op s); \
+    } \
+    __device__ __forceinline__ output_type ## 3 operator op(scalar_type s, const input_type ## 3 & b) \
+    { \
+        return VecTraits<output_type ## 3>::make(s op b.x, s op b.y, s op b.z); \
+    } \
+    __device__ __forceinline__ output_type ## 4 operator op(const input_type ## 4 & a, scalar_type s) \
+    { \
+        return VecTraits<output_type ## 4>::make(a.x op s, a.y op s, a.z op s, a.w op s); \
+    } \
+    __device__ __forceinline__ output_type ## 4 operator op(scalar_type s, const input_type ## 4 & b) \
+    { \
+        return VecTraits<output_type ## 4>::make(s op b.x, s op b.y, s op b.z, s op b.w); \
+    }
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uchar, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uchar, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uchar, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, char, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, char, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, char, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, ushort, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, ushort, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, ushort, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, short, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, short, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, short, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, int, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, int, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, int, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uint, uint, uint)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uint, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uint, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, float, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, float, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, double, double, double)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uchar, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uchar, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uchar, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, char, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, char, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, char, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, ushort, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, ushort, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, ushort, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, short, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, short, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, short, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, int, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, int, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, int, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uint, uint, uint)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uint, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uint, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, float, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, float, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, double, double, double)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uchar, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uchar, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uchar, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, char, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, char, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, char, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, ushort, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, ushort, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, ushort, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, short, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, short, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, short, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, int, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, int, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, int, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uint, uint, uint)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uint, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uint, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, float, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, float, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, double, double, double)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uchar, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uchar, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uchar, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, char, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, char, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, char, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, ushort, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, ushort, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, ushort, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, short, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, short, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, short, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, int, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, int, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, int, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uint, uint, uint)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uint, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uint, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, float, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, float, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, double, double, double)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, char, char, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, ushort, ushort, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, short, short, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, int, int, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, uint, uint, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, float, float, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, double, double, uchar)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, char, char, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, ushort, ushort, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, short, short, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, int, int, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, uint, uint, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, float, float, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, double, double, uchar)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, char, char, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, ushort, ushort, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, short, short, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, int, int, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, uint, uint, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, float, float, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, double, double, uchar)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, char, char, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, ushort, ushort, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, short, short, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, int, int, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, uint, uint, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, float, float, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, double, double, uchar)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, char, char, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, ushort, ushort, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, short, short, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, int, int, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, uint, uint, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, float, float, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, double, double, uchar)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, char, char, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, ushort, ushort, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, short, short, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, int, int, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, uint, uint, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, float, float, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, double, double, uchar)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, char, char, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, ushort, ushort, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, short, short, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, int, int, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, uint, uint, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, float, float, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, double, double, uchar)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, char, char, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, ushort, ushort, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, short, short, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, int, int, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, uint, uint, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, float, float, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, double, double, uchar)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, char, char, char)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, ushort, ushort, ushort)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, short, short, short)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, int, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, uint, uint, uint)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, char, char, char)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, ushort, ushort, ushort)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, short, short, short)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, int, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, uint, uint, uint)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, char, char, char)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, ushort, ushort, ushort)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, short, short, short)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, int, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, uint, uint, uint)
+
+#undef CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP
+
+// binary function (vec & vec)
+
+#define CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(func_name, func, input_type, output_type) \
+    __device__ __forceinline__ output_type ## 1 func_name(const input_type ## 1 & a, const input_type ## 1 & b) \
+    { \
+        return VecTraits<output_type ## 1>::make(func (a.x, b.x)); \
+    } \
+    __device__ __forceinline__ output_type ## 2 func_name(const input_type ## 2 & a, const input_type ## 2 & b) \
+    { \
+        return VecTraits<output_type ## 2>::make(func (a.x, b.x), func (a.y, b.y)); \
+    } \
+    __device__ __forceinline__ output_type ## 3 func_name(const input_type ## 3 & a, const input_type ## 3 & b) \
+    { \
+        return VecTraits<output_type ## 3>::make(func (a.x, b.x), func (a.y, b.y), func (a.z, b.z)); \
+    } \
+    __device__ __forceinline__ output_type ## 4 func_name(const input_type ## 4 & a, const input_type ## 4 & b) \
+    { \
+        return VecTraits<output_type ## 4>::make(func (a.x, b.x), func (a.y, b.y), func (a.z, b.z), func (a.w, b.w)); \
+    }
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, char, char)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, ushort, ushort)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, short, short)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, uint, uint)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, int, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::fmaxf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::fmax, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, uchar, uchar)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, char, char)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, ushort, ushort)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, short, short)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, uint, uint)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, int, int)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::fminf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::fmin, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, char, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, short, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, int, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, float, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypot, double, double)
+
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, uchar, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, char, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, ushort, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, short, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, uint, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, int, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, float, float)
+CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2, double, double)
+
+#undef CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC
+
+// binary function (vec & scalar)
+
+#define CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(func_name, func, input_type, scalar_type, output_type) \
+    __device__ __forceinline__ output_type ## 1 func_name(const input_type ## 1 & a, scalar_type s) \
+    { \
+        return VecTraits<output_type ## 1>::make(func ((output_type) a.x, (output_type) s)); \
+    } \
+    __device__ __forceinline__ output_type ## 1 func_name(scalar_type s, const input_type ## 1 & b) \
+    { \
+        return VecTraits<output_type ## 1>::make(func ((output_type) s, (output_type) b.x)); \
+    } \
+    __device__ __forceinline__ output_type ## 2 func_name(const input_type ## 2 & a, scalar_type s) \
+    { \
+        return VecTraits<output_type ## 2>::make(func ((output_type) a.x, (output_type) s), func ((output_type) a.y, (output_type) s)); \
+    } \
+    __device__ __forceinline__ output_type ## 2 func_name(scalar_type s, const input_type ## 2 & b) \
+    { \
+        return VecTraits<output_type ## 2>::make(func ((output_type) s, (output_type) b.x), func ((output_type) s, (output_type) b.y)); \
+    } \
+    __device__ __forceinline__ output_type ## 3 func_name(const input_type ## 3 & a, scalar_type s) \
+    { \
+        return VecTraits<output_type ## 3>::make(func ((output_type) a.x, (output_type) s), func ((output_type) a.y, (output_type) s), func ((output_type) a.z, (output_type) s)); \
+    } \
+    __device__ __forceinline__ output_type ## 3 func_name(scalar_type s, const input_type ## 3 & b) \
+    { \
+        return VecTraits<output_type ## 3>::make(func ((output_type) s, (output_type) b.x), func ((output_type) s, (output_type) b.y), func ((output_type) s, (output_type) b.z)); \
+    } \
+    __device__ __forceinline__ output_type ## 4 func_name(const input_type ## 4 & a, scalar_type s) \
+    { \
+        return VecTraits<output_type ## 4>::make(func ((output_type) a.x, (output_type) s), func ((output_type) a.y, (output_type) s), func ((output_type) a.z, (output_type) s), func ((output_type) a.w, (output_type) s)); \
+    } \
+    __device__ __forceinline__ output_type ## 4 func_name(scalar_type s, const input_type ## 4 & b) \
+    { \
+        return VecTraits<output_type ## 4>::make(func ((output_type) s, (output_type) b.x), func ((output_type) s, (output_type) b.y), func ((output_type) s, (output_type) b.z), func ((output_type) s, (output_type) b.w)); \
+    }
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, uchar, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, uchar, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, char, char, char)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, char, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, char, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, ushort, ushort, ushort)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, ushort, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, ushort, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, short, short, short)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, short, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, short, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, uint, uint, uint)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, uint, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, uint, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, int, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, int, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, int, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, float, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, float, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, double, double, double)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, uchar, uchar, uchar)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, uchar, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, uchar, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, char, char, char)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, char, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, char, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, ushort, ushort, ushort)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, ushort, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, ushort, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, short, short, short)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, short, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, short, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, uint, uint, uint)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, uint, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, uint, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, int, int, int)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, int, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, int, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, float, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, float, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, double, double, double)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, uchar, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, uchar, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, char, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, char, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, ushort, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, ushort, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, short, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, short, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, uint, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, uint, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, int, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, int, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, float, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, float, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, double, double, double)
+
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, uchar, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, uchar, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, char, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, char, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, ushort, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, ushort, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, short, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, short, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, uint, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, uint, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, int, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, int, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, float, float, float)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, float, double, double)
+CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, double, double, double)
+
+#undef CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC
+
+}}} // namespace cv { namespace cuda { namespace device
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_VECMATH_HPP__

3rdparty/include/opencv2/core/cuda/vec_traits.hpp

@@ -0,0 +1,288 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_VEC_TRAITS_HPP__
+#define __OPENCV_CUDA_VEC_TRAITS_HPP__
+
+#include "common.hpp"
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template<typename T, int N> struct TypeVec;
+
+    struct __align__(8) uchar8
+    {
+        uchar a0, a1, a2, a3, a4, a5, a6, a7;
+    };
+    static __host__ __device__ __forceinline__ uchar8 make_uchar8(uchar a0, uchar a1, uchar a2, uchar a3, uchar a4, uchar a5, uchar a6, uchar a7)
+    {
+        uchar8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
+        return val;
+    }
+    struct __align__(8) char8
+    {
+        schar a0, a1, a2, a3, a4, a5, a6, a7;
+    };
+    static __host__ __device__ __forceinline__ char8 make_char8(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7)
+    {
+        char8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
+        return val;
+    }
+    struct __align__(16) ushort8
+    {
+        ushort a0, a1, a2, a3, a4, a5, a6, a7;
+    };
+    static __host__ __device__ __forceinline__ ushort8 make_ushort8(ushort a0, ushort a1, ushort a2, ushort a3, ushort a4, ushort a5, ushort a6, ushort a7)
+    {
+        ushort8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
+        return val;
+    }
+    struct __align__(16) short8
+    {
+        short a0, a1, a2, a3, a4, a5, a6, a7;
+    };
+    static __host__ __device__ __forceinline__ short8 make_short8(short a0, short a1, short a2, short a3, short a4, short a5, short a6, short a7)
+    {
+        short8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
+        return val;
+    }
+    struct __align__(32) uint8
+    {
+        uint a0, a1, a2, a3, a4, a5, a6, a7;
+    };
+    static __host__ __device__ __forceinline__ uint8 make_uint8(uint a0, uint a1, uint a2, uint a3, uint a4, uint a5, uint a6, uint a7)
+    {
+        uint8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
+        return val;
+    }
+    struct __align__(32) int8
+    {
+        int a0, a1, a2, a3, a4, a5, a6, a7;
+    };
+    static __host__ __device__ __forceinline__ int8 make_int8(int a0, int a1, int a2, int a3, int a4, int a5, int a6, int a7)
+    {
+        int8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
+        return val;
+    }
+    struct __align__(32) float8
+    {
+        float a0, a1, a2, a3, a4, a5, a6, a7;
+    };
+    static __host__ __device__ __forceinline__ float8 make_float8(float a0, float a1, float a2, float a3, float a4, float a5, float a6, float a7)
+    {
+        float8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
+        return val;
+    }
+    struct double8
+    {
+        double a0, a1, a2, a3, a4, a5, a6, a7;
+    };
+    static __host__ __device__ __forceinline__ double8 make_double8(double a0, double a1, double a2, double a3, double a4, double a5, double a6, double a7)
+    {
+        double8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
+        return val;
+    }
+
+#define OPENCV_CUDA_IMPLEMENT_TYPE_VEC(type) \
+    template<> struct TypeVec<type, 1> { typedef type vec_type; }; \
+    template<> struct TypeVec<type ## 1, 1> { typedef type ## 1 vec_type; }; \
+    template<> struct TypeVec<type, 2> { typedef type ## 2 vec_type; }; \
+    template<> struct TypeVec<type ## 2, 2> { typedef type ## 2 vec_type; }; \
+    template<> struct TypeVec<type, 3> { typedef type ## 3 vec_type; }; \
+    template<> struct TypeVec<type ## 3, 3> { typedef type ## 3 vec_type; }; \
+    template<> struct TypeVec<type, 4> { typedef type ## 4 vec_type; }; \
+    template<> struct TypeVec<type ## 4, 4> { typedef type ## 4 vec_type; }; \
+    template<> struct TypeVec<type, 8> { typedef type ## 8 vec_type; }; \
+    template<> struct TypeVec<type ## 8, 8> { typedef type ## 8 vec_type; };
+
+    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(uchar)
+    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(char)
+    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(ushort)
+    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(short)
+    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(int)
+    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(uint)
+    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(float)
+    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(double)
+
+    #undef OPENCV_CUDA_IMPLEMENT_TYPE_VEC
+
+    template<> struct TypeVec<schar, 1> { typedef schar vec_type; };
+    template<> struct TypeVec<schar, 2> { typedef char2 vec_type; };
+    template<> struct TypeVec<schar, 3> { typedef char3 vec_type; };
+    template<> struct TypeVec<schar, 4> { typedef char4 vec_type; };
+    template<> struct TypeVec<schar, 8> { typedef char8 vec_type; };
+
+    template<> struct TypeVec<bool, 1> { typedef uchar vec_type; };
+    template<> struct TypeVec<bool, 2> { typedef uchar2 vec_type; };
+    template<> struct TypeVec<bool, 3> { typedef uchar3 vec_type; };
+    template<> struct TypeVec<bool, 4> { typedef uchar4 vec_type; };
+    template<> struct TypeVec<bool, 8> { typedef uchar8 vec_type; };
+
+    template<typename T> struct VecTraits;
+
+#define OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(type) \
+    template<> struct VecTraits<type> \
+    { \
+        typedef type elem_type; \
+        enum {cn=1}; \
+        static __device__ __host__ __forceinline__ type all(type v) {return v;} \
+        static __device__ __host__ __forceinline__ type make(type x) {return x;} \
+        static __device__ __host__ __forceinline__ type make(const type* v) {return *v;} \
+    }; \
+    template<> struct VecTraits<type ## 1> \
+    { \
+        typedef type elem_type; \
+        enum {cn=1}; \
+        static __device__ __host__ __forceinline__ type ## 1 all(type v) {return make_ ## type ## 1(v);} \
+        static __device__ __host__ __forceinline__ type ## 1 make(type x) {return make_ ## type ## 1(x);} \
+        static __device__ __host__ __forceinline__ type ## 1 make(const type* v) {return make_ ## type ## 1(*v);} \
+    }; \
+    template<> struct VecTraits<type ## 2> \
+    { \
+        typedef type elem_type; \
+        enum {cn=2}; \
+        static __device__ __host__ __forceinline__ type ## 2 all(type v) {return make_ ## type ## 2(v, v);} \
+        static __device__ __host__ __forceinline__ type ## 2 make(type x, type y) {return make_ ## type ## 2(x, y);} \
+        static __device__ __host__ __forceinline__ type ## 2 make(const type* v) {return make_ ## type ## 2(v[0], v[1]);} \
+    }; \
+    template<> struct VecTraits<type ## 3> \
+    { \
+        typedef type elem_type; \
+        enum {cn=3}; \
+        static __device__ __host__ __forceinline__ type ## 3 all(type v) {return make_ ## type ## 3(v, v, v);} \
+        static __device__ __host__ __forceinline__ type ## 3 make(type x, type y, type z) {return make_ ## type ## 3(x, y, z);} \
+        static __device__ __host__ __forceinline__ type ## 3 make(const type* v) {return make_ ## type ## 3(v[0], v[1], v[2]);} \
+    }; \
+    template<> struct VecTraits<type ## 4> \
+    { \
+        typedef type elem_type; \
+        enum {cn=4}; \
+        static __device__ __host__ __forceinline__ type ## 4 all(type v) {return make_ ## type ## 4(v, v, v, v);} \
+        static __device__ __host__ __forceinline__ type ## 4 make(type x, type y, type z, type w) {return make_ ## type ## 4(x, y, z, w);} \
+        static __device__ __host__ __forceinline__ type ## 4 make(const type* v) {return make_ ## type ## 4(v[0], v[1], v[2], v[3]);} \
+    }; \
+    template<> struct VecTraits<type ## 8> \
+    { \
+        typedef type elem_type; \
+        enum {cn=8}; \
+        static __device__ __host__ __forceinline__ type ## 8 all(type v) {return make_ ## type ## 8(v, v, v, v, v, v, v, v);} \
+        static __device__ __host__ __forceinline__ type ## 8 make(type a0, type a1, type a2, type a3, type a4, type a5, type a6, type a7) {return make_ ## type ## 8(a0, a1, a2, a3, a4, a5, a6, a7);} \
+        static __device__ __host__ __forceinline__ type ## 8 make(const type* v) {return make_ ## type ## 8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);} \
+    };
+
+    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(uchar)
+    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(ushort)
+    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(short)
+    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(int)
+    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(uint)
+    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(float)
+    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(double)
+
+    #undef OPENCV_CUDA_IMPLEMENT_VEC_TRAITS
+
+    template<> struct VecTraits<char>
+    {
+        typedef char elem_type;
+        enum {cn=1};
+        static __device__ __host__ __forceinline__ char all(char v) {return v;}
+        static __device__ __host__ __forceinline__ char make(char x) {return x;}
+        static __device__ __host__ __forceinline__ char make(const char* x) {return *x;}
+    };
+    template<> struct VecTraits<schar>
+    {
+        typedef schar elem_type;
+        enum {cn=1};
+        static __device__ __host__ __forceinline__ schar all(schar v) {return v;}
+        static __device__ __host__ __forceinline__ schar make(schar x) {return x;}
+        static __device__ __host__ __forceinline__ schar make(const schar* x) {return *x;}
+    };
+    template<> struct VecTraits<char1>
+    {
+        typedef schar elem_type;
+        enum {cn=1};
+        static __device__ __host__ __forceinline__ char1 all(schar v) {return make_char1(v);}
+        static __device__ __host__ __forceinline__ char1 make(schar x) {return make_char1(x);}
+        static __device__ __host__ __forceinline__ char1 make(const schar* v) {return make_char1(v[0]);}
+    };
+    template<> struct VecTraits<char2>
+    {
+        typedef schar elem_type;
+        enum {cn=2};
+        static __device__ __host__ __forceinline__ char2 all(schar v) {return make_char2(v, v);}
+        static __device__ __host__ __forceinline__ char2 make(schar x, schar y) {return make_char2(x, y);}
+        static __device__ __host__ __forceinline__ char2 make(const schar* v) {return make_char2(v[0], v[1]);}
+    };
+    template<> struct VecTraits<char3>
+    {
+        typedef schar elem_type;
+        enum {cn=3};
+        static __device__ __host__ __forceinline__ char3 all(schar v) {return make_char3(v, v, v);}
+        static __device__ __host__ __forceinline__ char3 make(schar x, schar y, schar z) {return make_char3(x, y, z);}
+        static __device__ __host__ __forceinline__ char3 make(const schar* v) {return make_char3(v[0], v[1], v[2]);}
+    };
+    template<> struct VecTraits<char4>
+    {
+        typedef schar elem_type;
+        enum {cn=4};
+        static __device__ __host__ __forceinline__ char4 all(schar v) {return make_char4(v, v, v, v);}
+        static __device__ __host__ __forceinline__ char4 make(schar x, schar y, schar z, schar w) {return make_char4(x, y, z, w);}
+        static __device__ __host__ __forceinline__ char4 make(const schar* v) {return make_char4(v[0], v[1], v[2], v[3]);}
+    };
+    template<> struct VecTraits<char8>
+    {
+        typedef schar elem_type;
+        enum {cn=8};
+        static __device__ __host__ __forceinline__ char8 all(schar v) {return make_char8(v, v, v, v, v, v, v, v);}
+        static __device__ __host__ __forceinline__ char8 make(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7) {return make_char8(a0, a1, a2, a3, a4, a5, a6, a7);}
+        static __device__ __host__ __forceinline__ char8 make(const schar* v) {return make_char8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);}
+    };
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_VEC_TRAITS_HPP__

3rdparty/include/opencv2/core/cuda/warp.hpp

@@ -0,0 +1,139 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_DEVICE_WARP_HPP__
+#define __OPENCV_CUDA_DEVICE_WARP_HPP__
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    struct Warp
+    {
+        enum
+        {
+            LOG_WARP_SIZE = 5,
+            WARP_SIZE     = 1 << LOG_WARP_SIZE,
+            STRIDE        = WARP_SIZE
+        };
+
+        /** \brief Returns the warp lane ID of the calling thread. */
+        static __device__ __forceinline__ unsigned int laneId()
+        {
+            unsigned int ret;
+            asm("mov.u32 %0, %laneid;" : "=r"(ret) );
+            return ret;
+        }
+
+        template<typename It, typename T>
+        static __device__ __forceinline__ void fill(It beg, It end, const T& value)
+        {
+            for(It t = beg + laneId(); t < end; t += STRIDE)
+                *t = value;
+        }
+
+        template<typename InIt, typename OutIt>
+        static __device__ __forceinline__ OutIt copy(InIt beg, InIt end, OutIt out)
+        {
+            for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
+                *out = *t;
+            return out;
+        }
+
+        template<typename InIt, typename OutIt, class UnOp>
+        static __device__ __forceinline__ OutIt transform(InIt beg, InIt end, OutIt out, UnOp op)
+        {
+            for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
+                *out = op(*t);
+            return out;
+        }
+
+        template<typename InIt1, typename InIt2, typename OutIt, class BinOp>
+        static __device__ __forceinline__ OutIt transform(InIt1 beg1, InIt1 end1, InIt2 beg2, OutIt out, BinOp op)
+        {
+            unsigned int lane = laneId();
+
+            InIt1 t1 = beg1 + lane;
+            InIt2 t2 = beg2 + lane;
+            for(; t1 < end1; t1 += STRIDE, t2 += STRIDE, out += STRIDE)
+                *out = op(*t1, *t2);
+            return out;
+        }
+
+        template <class T, class BinOp>
+        static __device__ __forceinline__ T reduce(volatile T *ptr, BinOp op)
+        {
+            const unsigned int lane = laneId();
+
+            if (lane < 16)
+            {
+                T partial = ptr[lane];
+
+                ptr[lane] = partial = op(partial, ptr[lane + 16]);
+                ptr[lane] = partial = op(partial, ptr[lane + 8]);
+                ptr[lane] = partial = op(partial, ptr[lane + 4]);
+                ptr[lane] = partial = op(partial, ptr[lane + 2]);
+                ptr[lane] = partial = op(partial, ptr[lane + 1]);
+            }
+
+            return *ptr;
+        }
+
+        template<typename OutIt, typename T>
+        static __device__ __forceinline__ void yota(OutIt beg, OutIt end, T value)
+        {
+            unsigned int lane = laneId();
+            value += lane;
+
+            for(OutIt t = beg + lane; t < end; t += STRIDE, value += STRIDE)
+                *t = value;
+        }
+    };
+}}} // namespace cv { namespace cuda { namespace cudev
+
+//! @endcond
+
+#endif /* __OPENCV_CUDA_DEVICE_WARP_HPP__ */

3rdparty/include/opencv2/core/cuda/warp_reduce.hpp

@@ -0,0 +1,76 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef OPENCV_CUDA_WARP_REDUCE_HPP__
+#define OPENCV_CUDA_WARP_REDUCE_HPP__
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template <class T>
+    __device__ __forceinline__ T warp_reduce(volatile T *ptr , const unsigned int tid = threadIdx.x)
+    {
+        const unsigned int lane = tid & 31; // index of thread in warp (0..31)
+
+        if (lane < 16)
+        {
+            T partial = ptr[tid];
+
+            ptr[tid] = partial = partial + ptr[tid + 16];
+            ptr[tid] = partial = partial + ptr[tid + 8];
+            ptr[tid] = partial = partial + ptr[tid + 4];
+            ptr[tid] = partial = partial + ptr[tid + 2];
+            ptr[tid] = partial = partial + ptr[tid + 1];
+        }
+
+        return ptr[tid - lane];
+    }
+}}} // namespace cv { namespace cuda { namespace cudev {
+
+//! @endcond
+
+#endif /* OPENCV_CUDA_WARP_REDUCE_HPP__ */

3rdparty/include/opencv2/core/cuda/warp_shuffle.hpp

@@ -0,0 +1,153 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_CUDA_WARP_SHUFFLE_HPP__
+#define __OPENCV_CUDA_WARP_SHUFFLE_HPP__
+
+/** @file
+ * @deprecated Use @ref cudev instead.
+ */
+
+//! @cond IGNORED
+
+namespace cv { namespace cuda { namespace device
+{
+    template <typename T>
+    __device__ __forceinline__ T shfl(T val, int srcLane, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return __shfl(val, srcLane, width);
+    #else
+        return T();
+    #endif
+    }
+    __device__ __forceinline__ unsigned int shfl(unsigned int val, int srcLane, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return (unsigned int) __shfl((int) val, srcLane, width);
+    #else
+        return 0;
+    #endif
+    }
+    __device__ __forceinline__ double shfl(double val, int srcLane, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        int lo = __double2loint(val);
+        int hi = __double2hiint(val);
+
+        lo = __shfl(lo, srcLane, width);
+        hi = __shfl(hi, srcLane, width);
+
+        return __hiloint2double(hi, lo);
+    #else
+        return 0.0;
+    #endif
+    }
+
+    template <typename T>
+    __device__ __forceinline__ T shfl_down(T val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return __shfl_down(val, delta, width);
+    #else
+        return T();
+    #endif
+    }
+    __device__ __forceinline__ unsigned int shfl_down(unsigned int val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return (unsigned int) __shfl_down((int) val, delta, width);
+    #else
+        return 0;
+    #endif
+    }
+    __device__ __forceinline__ double shfl_down(double val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        int lo = __double2loint(val);
+        int hi = __double2hiint(val);
+
+        lo = __shfl_down(lo, delta, width);
+        hi = __shfl_down(hi, delta, width);
+
+        return __hiloint2double(hi, lo);
+    #else
+        return 0.0;
+    #endif
+    }
+
+    template <typename T>
+    __device__ __forceinline__ T shfl_up(T val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return __shfl_up(val, delta, width);
+    #else
+        return T();
+    #endif
+    }
+    __device__ __forceinline__ unsigned int shfl_up(unsigned int val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return (unsigned int) __shfl_up((int) val, delta, width);
+    #else
+        return 0;
+    #endif
+    }
+    __device__ __forceinline__ double shfl_up(double val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        int lo = __double2loint(val);
+        int hi = __double2hiint(val);
+
+        lo = __shfl_up(lo, delta, width);
+        hi = __shfl_up(hi, delta, width);
+
+        return __hiloint2double(hi, lo);
+    #else
+        return 0.0;
+    #endif
+    }
+}}}
+
+//! @endcond
+
+#endif // __OPENCV_CUDA_WARP_SHUFFLE_HPP__