Merge pull request #1903 from ilya-lavrenov:tapi_warp

modules/core/include/opencv2/core/ocl.hpp

@@ -286,7 +286,7 @@ public:
     Kernel();
     Kernel(const char* kname, const Program& prog);
     Kernel(const char* kname, const ProgramSource2& prog,
-           const String& buildopts, String* errmsg=0);
+           const String& buildopts = String(), String* errmsg=0);
     ~Kernel();
     Kernel(const Kernel& k);
     Kernel& operator = (const Kernel& k);

modules/core/src/umatrix.cpp

@@ -578,6 +578,7 @@ Mat UMat::getMat(int accessFlags) const
     u->currAllocator->map(u, accessFlags | ACCESS_READ);
     CV_Assert(u->data != 0);
     Mat hdr(dims, size.p, type(), u->data + offset, step.p);
+    hdr.flags = flags;
     hdr.u = u;
     hdr.datastart = u->data;
     hdr.data = hdr.datastart + offset;

modules/imgproc/src/imgwarp.cpp

@@ -2010,6 +2010,8 @@ static bool ocl_resize( InputArray _src, OutputArray _dst, Size dsize,
                                   iscale_x, iscale_y, 1.0f / (iscale_x * iscale_y));
 
             k.create("resizeAREA_FAST", ocl::imgproc::resize_oclsrc, buildOption);
+            if (k.empty())
+                return false;
 
             int smap_tab_size = dst.cols * iscale_x + dst.rows * iscale_y;
             AutoBuffer<int> dmap_tab(dst.cols + dst.rows), smap_tab(smap_tab_size);
@@ -2026,6 +2028,8 @@ static bool ocl_resize( InputArray _src, OutputArray _dst, Size dsize,
         {
             buildOption = buildOption + format(" -D convertToT=%s", ocl::convertTypeStr(wdepth, depth, cn, cvt[0]));
             k.create("resizeAREA", ocl::imgproc::resize_oclsrc, buildOption);
+            if (k.empty())
+                return false;
 
             Size ssize = src.size();
             int xytab_size = (ssize.width + ssize.height) << 1;
@@ -3383,6 +3387,78 @@ private:
     const void *ctab;
 };
 
+static bool ocl_remap(InputArray _src, OutputArray _dst, InputArray _map1, InputArray _map2,
+                      int interpolation, int borderType, const Scalar& borderValue)
+{
+    int cn = _src.channels(), type = _src.type(), depth = _src.depth();
+
+    if (borderType == BORDER_TRANSPARENT || cn == 3 || !(interpolation == INTER_LINEAR || interpolation == INTER_NEAREST)
+            || _map1.type() == CV_16SC1 || _map2.type() == CV_16SC1)
+        return false;
+
+    UMat src = _src.getUMat(), map1 = _map1.getUMat(), map2 = _map2.getUMat();
+
+    if( (map1.type() == CV_16SC2 && (map2.type() == CV_16UC1 || map2.empty())) ||
+        (map2.type() == CV_16SC2 && (map1.type() == CV_16UC1 || map1.empty())) )
+    {
+        if (map1.type() != CV_16SC2)
+            std::swap(map1, map2);
+    }
+    else
+        CV_Assert( map1.type() == CV_32FC2 || (map1.type() == CV_32FC1 && map2.type() == CV_32FC1) );
+
+    _dst.create(map1.size(), type);
+    UMat dst = _dst.getUMat();
+
+    String kernelName = "remap";
+    if (map1.type() == CV_32FC2 && map2.empty())
+        kernelName += "_32FC2";
+    else if (map1.type() == CV_16SC2)
+    {
+        kernelName += "_16SC2";
+        if (!map2.empty())
+            kernelName += "_16UC1";
+    }
+    else if (map1.type() == CV_32FC1 && map2.type() == CV_32FC1)
+        kernelName += "_2_32FC1";
+    else
+        CV_Error(Error::StsBadArg, "Unsupported map types");
+
+    static const char * const interMap[] = { "INTER_NEAREST", "INTER_LINEAR", "INTER_CUBIC", "INTER_LINEAR", "INTER_LANCZOS" };
+    static const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", "BORDER_WRAP",
+                           "BORDER_REFLECT_101", "BORDER_TRANSPARENT" };
+    String buildOptions = format("-D %s -D %s -D T=%s", interMap[interpolation], borderMap[borderType], ocl::typeToStr(type));
+
+    if (interpolation != INTER_NEAREST)
+    {
+        char cvt[3][40];
+        int wdepth = std::max(CV_32F, dst.depth());
+        buildOptions = buildOptions
+                      + format(" -D WT=%s -D convertToT=%s -D convertToWT=%s"
+                               " -D convertToWT2=%s -D WT2=%s",
+                               ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
+                               ocl::convertTypeStr(wdepth, depth, cn, cvt[0]),
+                               ocl::convertTypeStr(depth, wdepth, cn, cvt[1]),
+                               ocl::convertTypeStr(CV_32S, wdepth, 2, cvt[2]),
+                               ocl::typeToStr(CV_MAKE_TYPE(wdepth, 2)));
+    }
+
+    ocl::Kernel k(kernelName.c_str(), ocl::imgproc::remap_oclsrc, buildOptions);
+
+    Mat scalar(1, 1, type, borderValue);
+    ocl::KernelArg srcarg = ocl::KernelArg::ReadOnly(src), dstarg = ocl::KernelArg::WriteOnly(dst),
+            map1arg = ocl::KernelArg::ReadOnlyNoSize(map1),
+            scalararg = ocl::KernelArg::Constant((void*)scalar.data, scalar.elemSize());
+
+    if (map2.empty())
+        k.args(srcarg, dstarg, map1arg, scalararg);
+    else
+        k.args(srcarg, dstarg, map1arg, ocl::KernelArg::ReadOnlyNoSize(map2), scalararg);
+
+    size_t globalThreads[2] = { dst.cols, dst.rows };
+    return k.run(2, globalThreads, NULL, false);
+}
+
 }
 
 void cv::remap( InputArray _src, OutputArray _dst,
@@ -3422,11 +3498,13 @@ void cv::remap( InputArray _src, OutputArray _dst,
         remapLanczos4<Cast<double, double>, float, 1>, 0
     };
 
-    Mat src = _src.getMat(), map1 = _map1.getMat(), map2 = _map2.getMat();
+    CV_Assert( _map1.size().area() > 0 );
+    CV_Assert( _map2.empty() || (_map2.size() == _map1.size()));
 
-    CV_Assert( map1.size().area() > 0 );
-    CV_Assert( !map2.data || (map2.size() == map1.size()));
+    if (ocl::useOpenCL() && _dst.isUMat() && ocl_remap(_src, _dst, _map1, _map2, interpolation, borderType, borderValue))
+        return;
 
+    Mat src = _src.getMat(), map1 = _map1.getMat(), map2 = _map2.getMat();
     _dst.create( map1.size(), src.type() );
     Mat dst = _dst.getMat();
     if( dst.data == src.data )
@@ -3789,6 +3867,89 @@ private:
 };
 #endif
 
+enum { OCL_OP_PERSPECTIVE = 1, OCL_OP_AFFINE = 0 };
+
+static bool ocl_warpTransform(InputArray _src, OutputArray _dst, InputArray _M0,
+                              Size dsize, int flags, int borderType, const Scalar& borderValue,
+                              int op_type)
+{
+    CV_Assert(op_type == OCL_OP_AFFINE || op_type == OCL_OP_PERSPECTIVE);
+
+    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), wdepth = depth;
+    double doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
+
+    int interpolation = flags & INTER_MAX;
+    if( interpolation == INTER_AREA )
+        interpolation = INTER_LINEAR;
+
+    if ( !(borderType == cv::BORDER_CONSTANT &&
+           (interpolation == cv::INTER_NEAREST || interpolation == cv::INTER_LINEAR || interpolation == cv::INTER_CUBIC)) ||
+         (!doubleSupport && depth == CV_64F) || cn > 4 || cn == 3)
+        return false;
+
+    const char * const interpolationMap[3] = { "NEAREST", "LINEAR", "CUBIC" };
+    ocl::ProgramSource2 program = op_type == OCL_OP_AFFINE ?
+                ocl::imgproc::warp_affine_oclsrc : ocl::imgproc::warp_perspective_oclsrc;
+    const char * const kernelName = op_type == OCL_OP_AFFINE ? "warpAffine" : "warpPerspective";
+
+    ocl::Kernel k;
+    if (interpolation == INTER_NEAREST)
+    {
+        k.create(kernelName, program,
+                 format("-D INTER_NEAREST -D T=%s%s", ocl::typeToStr(type),
+                        doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
+    }
+    else
+    {
+        char cvt[2][50];
+        wdepth = std::max(CV_32S, depth);
+        k.create(kernelName, program,
+                  format("-D INTER_%s -D T=%s -D WT=%s -D depth=%d -D convertToWT=%s -D convertToT=%s%s",
+                         interpolationMap[interpolation], ocl::typeToStr(type),
+                         ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)), depth,
+                         ocl::convertTypeStr(depth, wdepth, cn, cvt[0]),
+                         ocl::convertTypeStr(wdepth, depth, cn, cvt[1]),
+                         doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
+    }
+    if (k.empty())
+        return false;
+
+    UMat src = _src.getUMat(), M0;
+    _dst.create( dsize.area() == 0 ? src.size() : dsize, src.type() );
+    UMat dst = _dst.getUMat();
+
+    double M[9];
+    int matRows = (op_type == OCL_OP_AFFINE ? 2 : 3);
+    Mat matM(matRows, 3, CV_64F, M), M1 = _M0.getMat();
+    CV_Assert( (M1.type() == CV_32F || M1.type() == CV_64F) &&
+               M1.rows == matRows && M1.cols == 3 );
+    M1.convertTo(matM, matM.type());
+
+    if( !(flags & WARP_INVERSE_MAP) )
+    {
+        if (op_type == OCL_OP_PERSPECTIVE)
+            invert(matM, matM);
+        else
+        {
+            double D = M[0]*M[4] - M[1]*M[3];
+            D = D != 0 ? 1./D : 0;
+            double A11 = M[4]*D, A22=M[0]*D;
+            M[0] = A11; M[1] *= -D;
+            M[3] *= -D; M[4] = A22;
+            double b1 = -M[0]*M[2] - M[1]*M[5];
+            double b2 = -M[3]*M[2] - M[4]*M[5];
+            M[2] = b1; M[5] = b2;
+        }
+    }
+    matM.convertTo(M0, doubleSupport ? CV_64F : CV_32F);
+
+    k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst), ocl::KernelArg::PtrReadOnly(M0),
+           ocl::KernelArg::Constant(Mat(1, 1, CV_MAKE_TYPE(wdepth, cn), borderValue)));
+
+    size_t globalThreads[2] = { dst.cols, dst.rows };
+    return k.run(2, globalThreads, NULL, false);
+}
+
 }
 
 
@@ -3796,6 +3957,11 @@ void cv::warpAffine( InputArray _src, OutputArray _dst,
                      InputArray _M0, Size dsize,
                      int flags, int borderType, const Scalar& borderValue )
 {
+    if (ocl::useOpenCL() && _dst.isUMat() &&
+        ocl_warpTransform(_src, _dst, _M0, dsize, flags, borderType,
+                                            borderValue, OCL_OP_AFFINE))
+        return;
+
     Mat src = _src.getMat(), M0 = _M0.getMat();
     _dst.create( dsize.area() == 0 ? src.size() : dsize, src.type() );
     Mat dst = _dst.getMat();
@@ -4035,11 +4201,17 @@ private:
 void cv::warpPerspective( InputArray _src, OutputArray _dst, InputArray _M0,
                           Size dsize, int flags, int borderType, const Scalar& borderValue )
 {
+    CV_Assert( _src.total() > 0 );
+
+    if (ocl::useOpenCL() && _dst.isUMat() &&
+            ocl_warpTransform(_src, _dst, _M0, dsize, flags, borderType, borderValue,
+                              OCL_OP_PERSPECTIVE))
+        return;
+
     Mat src = _src.getMat(), M0 = _M0.getMat();
     _dst.create( dsize.area() == 0 ? src.size() : dsize, src.type() );
     Mat dst = _dst.getMat();
 
-    CV_Assert( src.cols > 0 && src.rows > 0 );
     if( dst.data == src.data )
         src = src.clone();
 

modules/imgproc/src/opencl/warp_affine.cl

+/*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) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
+// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// @Authors
+//    Zhang Ying, zhangying913@gmail.com
+//
+// 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*/
+
+#ifdef DOUBLE_SUPPORT
+#ifdef cl_amd_fp64
+#pragma OPENCL EXTENSION cl_amd_fp64:enable
+#elif defined (cl_khr_fp64)
+#pragma OPENCL EXTENSION cl_khr_fp64:enable
+#endif
+#define CT double
+#else
+#define CT float
+#endif
+
+#define INTER_BITS 5
+#define INTER_TAB_SIZE (1 << INTER_BITS)
+#define INTER_SCALE 1.f/INTER_TAB_SIZE
+#define AB_BITS max(10, (int)INTER_BITS)
+#define AB_SCALE (1 << AB_BITS)
+#define INTER_REMAP_COEF_BITS 15
+#define INTER_REMAP_COEF_SCALE (1 << INTER_REMAP_COEF_BITS)
+
+#define noconvert
+
+#ifdef INTER_NEAREST
+
+__kernel void warpAffine(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
+                         __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols,
+                         __constant CT * M, T scalar)
+{
+    int dx = get_global_id(0);
+    int dy = get_global_id(1);
+
+    if (dx < dst_cols && dy < dst_rows)
+    {
+        int round_delta = (AB_SCALE >> 1);
+
+        int X0 = rint(M[0] * dx * AB_SCALE);
+        int Y0 = rint(M[3] * dx * AB_SCALE);
+        X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta;
+        Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta;
+
+        short sx = convert_short_sat(X0 >> AB_BITS);
+        short sy = convert_short_sat(Y0 >> AB_BITS);
+
+        int dst_index = mad24(dy, dst_step, dst_offset + dx * (int)sizeof(T));
+        __global T * dst = (__global T *)(dstptr + dst_index);
+
+        if (sx >= 0 && sx < src_cols && sy >= 0 && sy < src_rows)
+        {
+            int src_index = mad24(sy, src_step, src_offset + sx * (int)sizeof(T));
+            __global const T * src = (__global const T *)(srcptr + src_index);
+            dst[0] = src[0];
+        }
+        else
+            dst[0] = scalar;
+    }
+}
+
+#elif defined INTER_LINEAR
+
+__kernel void warpAffine(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
+                         __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols,
+                         __constant CT * M, WT scalar)
+{
+    int dx = get_global_id(0);
+    int dy = get_global_id(1);
+
+    if (dx < dst_cols && dy < dst_rows)
+    {
+        int round_delta = AB_SCALE/INTER_TAB_SIZE/2;
+
+        int tmp = (dx << AB_BITS);
+        int X0 = rint(M[0] * tmp);
+        int Y0 = rint(M[3] * tmp);
+        X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta;
+        Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta;
+        X0 = X0 >> (AB_BITS - INTER_BITS);
+        Y0 = Y0 >> (AB_BITS - INTER_BITS);
+
+        short sx = convert_short_sat(X0 >> INTER_BITS);
+        short sy = convert_short_sat(Y0 >> INTER_BITS);
+        short ax = convert_short(X0 & (INTER_TAB_SIZE-1));
+        short ay = convert_short(Y0 & (INTER_TAB_SIZE-1));
+
+        WT v0 = (sx >= 0 && sx < src_cols && sy >= 0 && sy < src_rows) ?
+            convertToWT(*(__global const T *)(srcptr + mad24(sy, src_step, src_offset + sx * (int)sizeof(T)))) : scalar;
+        WT v1 = (sx+1 >= 0 && sx+1 < src_cols && sy >= 0 && sy < src_rows) ?
+            convertToWT(*(__global const T *)(srcptr + mad24(sy, src_step, src_offset + (sx+1) * (int)sizeof(T)))) : scalar;
+        WT v2 = (sx >= 0 && sx < src_cols && sy+1 >= 0 && sy+1 < src_rows) ?
+            convertToWT(*(__global const T *)(srcptr + mad24(sy+1, src_step, src_offset + sx * (int)sizeof(T)))) : scalar;
+        WT v3 = (sx+1 >= 0 && sx+1 < src_cols && sy+1 >= 0 && sy+1 < src_rows) ?
+            convertToWT(*(__global const T *)(srcptr + mad24(sy+1, src_step, src_offset + (sx+1) * (int)sizeof(T)))) : scalar;
+
+        float taby = 1.f/INTER_TAB_SIZE*ay;
+        float tabx = 1.f/INTER_TAB_SIZE*ax;
+
+        int dst_index = mad24(dy, dst_step, dst_offset + dx * (int)sizeof(T));
+        __global T * dst = (__global T *)(dstptr + dst_index);
+
+#if depth <= 4
+        int itab0 = convert_short_sat_rte( (1.0f-taby)*(1.0f-tabx) * INTER_REMAP_COEF_SCALE );
+        int itab1 = convert_short_sat_rte( (1.0f-taby)*tabx * INTER_REMAP_COEF_SCALE );
+        int itab2 = convert_short_sat_rte( taby*(1.0f-tabx) * INTER_REMAP_COEF_SCALE );
+        int itab3 = convert_short_sat_rte( taby*tabx * INTER_REMAP_COEF_SCALE );
+
+        WT val = v0 * itab0 +  v1 * itab1 + v2 * itab2 + v3 * itab3;
+        dst[0] = convertToT((val + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS);
+#else
+        float tabx2 = 1.0f - tabx, taby2 = 1.0f - taby;
+        WT val = v0 * tabx2 * taby2 +  v1 * tabx * taby2 + v2 * tabx2 * taby + v3 * tabx * taby;
+        dst[0] = convertToT(val);
+#endif
+    }
+}
+
+#elif defined INTER_CUBIC
+
+inline void interpolateCubic( float x, float* coeffs )
+{
+    const float A = -0.75f;
+
+    coeffs[0] = ((A*(x + 1.f) - 5.0f*A)*(x + 1.f) + 8.0f*A)*(x + 1.f) - 4.0f*A;
+    coeffs[1] = ((A + 2.f)*x - (A + 3.f))*x*x + 1.f;
+    coeffs[2] = ((A + 2.f)*(1.f - x) - (A + 3.f))*(1.f - x)*(1.f - x) + 1.f;
+    coeffs[3] = 1.f - coeffs[0] - coeffs[1] - coeffs[2];
+}
+
+__kernel void warpAffine(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
+                         __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols,
+                         __constant CT * M, WT scalar)
+{
+    int dx = get_global_id(0);
+    int dy = get_global_id(1);
+
+    if (dx < dst_cols && dy < dst_rows)
+    {
+        int round_delta = ((AB_SCALE>>INTER_BITS)>>1);
+
+        int tmp = (dx << AB_BITS);
+        int X0 = rint(M[0] * tmp);
+        int Y0 = rint(M[3] * tmp);
+        X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta;
+        Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta;
+        X0 = X0 >> (AB_BITS - INTER_BITS);
+        Y0 = Y0 >> (AB_BITS - INTER_BITS);
+
+        int sx = (short)(X0 >> INTER_BITS) - 1;
+        int sy = (short)(Y0 >> INTER_BITS) - 1;
+        int ay = (short)(Y0 & (INTER_TAB_SIZE-1));
+        int ax = (short)(X0 & (INTER_TAB_SIZE-1));
+
+        WT v[16];
+        #pragma unroll
+        for (int y = 0; y < 4; y++)
+            #pragma unroll
+            for (int x = 0; x < 4; x++)
+                v[mad24(y, 4, x)] = (sx+x >= 0 && sx+x < src_cols && sy+y >= 0 && sy+y < src_rows) ?
+                    convertToWT(*(__global const T *)(srcptr + mad24(sy+y, src_step, src_offset + (sx+x) * (int)sizeof(T)))) : scalar;
+
+        float tab1y[4], tab1x[4];
+
+        float ayy = INTER_SCALE * ay;
+        float axx = INTER_SCALE * ax;
+        interpolateCubic(ayy, tab1y);
+        interpolateCubic(axx, tab1x);
+
+        int dst_index = mad24(dy, dst_step, dst_offset + dx * (int)sizeof(T));
+        __global T * dst = (__global T *)(dstptr + dst_index);
+
+        WT sum = (WT)(0);
+#if depth <= 4
+        int itab[16];
+
+        #pragma unroll
+        for (int i = 0; i < 16; i++)
+            itab[i] = rint(tab1y[(i>>2)] * tab1x[(i&3)] * INTER_REMAP_COEF_SCALE);
+
+        #pragma unroll
+        for (int i = 0; i < 16; i++)
+            sum += v[i] * itab[i];
+        dst[0] = convertToT( (sum + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS );
+#else
+        #pragma unroll
+        for (int i = 0; i < 16; i++)
+            sum += v[i] * tab1y[(i>>2)] * tab1x[(i&3)];
+        dst[0] = convertToT( sum );
+#endif
+    }
+}
+
+#endif

modules/imgproc/src/opencl/warp_perspective.cl

+/*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) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
+// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// @Authors
+//    Zhang Ying, zhangying913@gmail.com
+//
+// 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*/
+
+#ifdef DOUBLE_SUPPORT
+#ifdef cl_amd_fp64
+#pragma OPENCL EXTENSION cl_amd_fp64:enable
+#elif defined (cl_khr_fp64)
+#pragma OPENCL EXTENSION cl_khr_fp64:enable
+#endif
+#define CT double
+#else
+#define CT float
+#endif
+
+#define INTER_BITS 5
+#define INTER_TAB_SIZE (1 << INTER_BITS)
+#define INTER_SCALE 1.f / INTER_TAB_SIZE
+#define AB_BITS max(10, (int)INTER_BITS)
+#define AB_SCALE (1 << AB_BITS)
+#define INTER_REMAP_COEF_BITS 15
+#define INTER_REMAP_COEF_SCALE (1 << INTER_REMAP_COEF_BITS)
+
+#define noconvert
+
+#ifdef INTER_NEAREST
+
+__kernel void warpPerspective(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
+                              __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols,
+                              __constant CT * M, T scalar)
+{
+    int dx = get_global_id(0);
+    int dy = get_global_id(1);
+
+    if (dx < dst_cols && dy < dst_rows)
+    {
+        CT X0 = M[0] * dx + M[1] * dy + M[2];
+        CT Y0 = M[3] * dx + M[4] * dy + M[5];
+        CT W = M[6] * dx + M[7] * dy + M[8];
+        W = W != 0.0f ? 1.f / W : 0.0f;
+        short sx = convert_short_sat_rte(X0*W);
+        short sy = convert_short_sat_rte(Y0*W);
+
+        int dst_index = mad24(dy, dst_step, dx * (int)sizeof(T) + dst_offset);
+        __global T * dst = (__global T *)(dstptr + dst_index);
+
+        if (sx >= 0 && sx < src_cols && sy >= 0 && sy < src_rows)
+        {
+            int src_index = mad24(sy, src_step, sx * (int)sizeof(T) + src_offset);
+            __global const T * src = (__global const T *)(srcptr + src_index);
+            dst[0] = src[0];
+        }
+        else
+            dst[0] = scalar;
+    }
+}
+
+#elif defined INTER_LINEAR
+
+__kernel void warpPerspective(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
+                              __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols,
+                              __constant CT * M, WT scalar)
+{
+    int dx = get_global_id(0);
+    int dy = get_global_id(1);
+
+    if (dx < dst_cols && dy < dst_rows)
+    {
+        CT X0 = M[0] * dx + M[1] * dy + M[2];
+        CT Y0 = M[3] * dx + M[4] * dy + M[5];
+        CT W = M[6] * dx + M[7] * dy + M[8];
+        W = W != 0.0f ? INTER_TAB_SIZE / W : 0.0f;
+        int X = rint(X0 * W), Y = rint(Y0 * W);
+
+        short sx = convert_short_sat(X >> INTER_BITS);
+        short sy = convert_short_sat(Y >> INTER_BITS);
+        short ay = (short)(Y & (INTER_TAB_SIZE - 1));
+        short ax = (short)(X & (INTER_TAB_SIZE - 1));
+
+        WT v0 = (sx >= 0 && sx < src_cols && sy >= 0 && sy < src_rows) ?
+            convertToWT(*(__global const T *)(srcptr + mad24(sy, src_step, src_offset + sx * (int)sizeof(T)))) : scalar;
+        WT v1 = (sx+1 >= 0 && sx+1 < src_cols && sy >= 0 && sy < src_rows) ?
+            convertToWT(*(__global const T *)(srcptr + mad24(sy, src_step, src_offset + (sx+1) * (int)sizeof(T)))) : scalar;
+        WT v2 = (sx >= 0 && sx < src_cols && sy+1 >= 0 && sy+1 < src_rows) ?
+            convertToWT(*(__global const T *)(srcptr + mad24(sy+1, src_step, src_offset + sx * (int)sizeof(T)))) : scalar;
+        WT v3 = (sx+1 >= 0 && sx+1 < src_cols && sy+1 >= 0 && sy+1 < src_rows) ?
+            convertToWT(*(__global const T *)(srcptr + mad24(sy+1, src_step, src_offset + (sx+1) * (int)sizeof(T)))) : scalar;
+
+        float taby = 1.f/INTER_TAB_SIZE*ay;
+        float tabx = 1.f/INTER_TAB_SIZE*ax;
+
+        int dst_index = mad24(dy, dst_step, dst_offset + dx * (int)sizeof(T));
+        __global T * dst = (__global T *)(dstptr + dst_index);
+
+#if depth <= 4
+        int itab0 = convert_short_sat_rte( (1.0f-taby)*(1.0f-tabx) * INTER_REMAP_COEF_SCALE );
+        int itab1 = convert_short_sat_rte( (1.0f-taby)*tabx * INTER_REMAP_COEF_SCALE );
+        int itab2 = convert_short_sat_rte( taby*(1.0f-tabx) * INTER_REMAP_COEF_SCALE );
+        int itab3 = convert_short_sat_rte( taby*tabx * INTER_REMAP_COEF_SCALE );
+
+        WT val = v0 * itab0 +  v1 * itab1 + v2 * itab2 + v3 * itab3;
+        dst[0] = convertToT((val + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS);
+#else
+        float tabx2 = 1.0f - tabx, taby2 = 1.0f - taby;
+        WT val = v0 * tabx2 * taby2 +  v1 * tabx * taby2 + v2 * tabx2 * taby + v3 * tabx * taby;
+        dst[0] = convertToT(val);
+#endif
+    }
+}
+
+#elif defined INTER_CUBIC
+
+inline void interpolateCubic( float x, float* coeffs )
+{
+    const float A = -0.75f;
+
+    coeffs[0] = ((A*(x + 1.f) - 5.0f*A)*(x + 1.f) + 8.0f*A)*(x + 1.f) - 4.0f*A;
+    coeffs[1] = ((A + 2.f)*x - (A + 3.f))*x*x + 1.f;
+    coeffs[2] = ((A + 2.f)*(1.f - x) - (A + 3.f))*(1.f - x)*(1.f - x) + 1.f;
+    coeffs[3] = 1.f - coeffs[0] - coeffs[1] - coeffs[2];
+}
+
+__kernel void warpPerspective(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
+                              __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols,
+                              __constant CT * M, WT scalar)
+{
+    int dx = get_global_id(0);
+    int dy = get_global_id(1);
+
+    if (dx < dst_cols && dy < dst_rows)
+    {
+        CT X0 = M[0] * dx + M[1] * dy + M[2];
+        CT Y0 = M[3] * dx + M[4] * dy + M[5];
+        CT W = M[6] * dx + M[7] * dy + M[8];
+        W = W != 0.0f ? INTER_TAB_SIZE / W : 0.0f;
+        int X = rint(X0 * W), Y = rint(Y0 * W);
+
+        short sx = convert_short_sat(X >> INTER_BITS) - 1;
+        short sy = convert_short_sat(Y >> INTER_BITS) - 1;
+        short ay = (short)(Y & (INTER_TAB_SIZE-1));
+        short ax = (short)(X & (INTER_TAB_SIZE-1));
+
+        WT v[16];
+        #pragma unroll
+        for (int y = 0; y < 4; y++)
+            #pragma unroll
+            for (int x = 0; x < 4; x++)
+                v[mad24(y, 4, x)] = (sx+x >= 0 && sx+x < src_cols && sy+y >= 0 && sy+y < src_rows) ?
+                    convertToWT(*(__global const T *)(srcptr + mad24(sy+y, src_step, src_offset + (sx+x) * (int)sizeof(T)))) : scalar;
+
+        float tab1y[4], tab1x[4];
+
+        float ayy = INTER_SCALE * ay;
+        float axx = INTER_SCALE * ax;
+        interpolateCubic(ayy, tab1y);
+        interpolateCubic(axx, tab1x);
+
+        int dst_index = mad24(dy, dst_step, dst_offset + dx * (int)sizeof(T));
+        __global T * dst = (__global T *)(dstptr + dst_index);
+
+        WT sum = (WT)(0);
+#if depth <= 4
+        int itab[16];
+
+        #pragma unroll
+        for (int i = 0; i < 16; i++)
+            itab[i] = rint(tab1y[(i>>2)] * tab1x[(i&3)] * INTER_REMAP_COEF_SCALE);
+
+        #pragma unroll
+        for (int i = 0; i < 16; i++)
+            sum += v[i] * itab[i];
+        dst[0] = convertToT( (sum + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS );
+#else
+        #pragma unroll
+        for (int i = 0; i < 16; i++)
+            sum += v[i] * tab1y[(i>>2)] * tab1x[(i&3)];
+        dst[0] = convertToT( sum );
+#endif
+    }
+}
+
+#endif

modules/imgproc/test/ocl/test_warp.cpp

@@ -60,8 +60,105 @@
 namespace cvtest {
 namespace ocl {
 
+enum
+{
+    noType = -1
+};
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+// warpAffine  & warpPerspective
+
+PARAM_TEST_CASE(WarpTestBase, MatType, Interpolation, bool, bool)
+{
+    int type, interpolation;
+    Size dsize;
+    bool useRoi, mapInverse;
+
+    TEST_DECLARE_INPUT_PARAMETER(src)
+    TEST_DECLARE_OUTPUT_PARAMETER(dst)
+
+    virtual void SetUp()
+    {
+        type = GET_PARAM(0);
+        interpolation = GET_PARAM(1);
+        mapInverse = GET_PARAM(2);
+        useRoi = GET_PARAM(3);
+
+        if (mapInverse)
+            interpolation |= WARP_INVERSE_MAP;
+    }
+
+    void random_roi()
+    {
+        dsize = randomSize(1, MAX_VALUE);
+
+        Size roiSize = randomSize(1, MAX_VALUE);
+        Border srcBorder = randomBorder(0, useRoi ? MAX_VALUE : 0);
+        randomSubMat(src, src_roi, roiSize, srcBorder, type, -MAX_VALUE, MAX_VALUE);
+
+        Border dstBorder = randomBorder(0, useRoi ? MAX_VALUE : 0);
+        randomSubMat(dst, dst_roi, dsize, dstBorder, type, -MAX_VALUE, MAX_VALUE);
+
+        UMAT_UPLOAD_INPUT_PARAMETER(src)
+        UMAT_UPLOAD_OUTPUT_PARAMETER(dst)
+    }
+
+    void Near(double threshold = 0.0)
+    {
+        EXPECT_MAT_NEAR(dst, udst, threshold);
+        EXPECT_MAT_NEAR(dst_roi, udst_roi, threshold);
+    }
+};
+
+/////warpAffine
+
+typedef WarpTestBase WarpAffine;
+
+OCL_TEST_P(WarpAffine, Mat)
+{
+    for (int j = 0; j < test_loop_times; j++)
+    {
+        random_roi();
+
+        Mat M = getRotationMatrix2D(Point2f(src_roi.cols / 2.0f, src_roi.rows / 2.0f),
+            rng.uniform(-180.f, 180.f), rng.uniform(0.4f, 2.0f));
+
+        OCL_OFF(cv::warpAffine(src_roi, dst_roi, M, dsize, interpolation));
+        OCL_ON(cv::warpAffine(usrc_roi, udst_roi, M, dsize, interpolation));
+
+        Near(1.0);
+    }
+}
+
+//// warpPerspective
+
+typedef WarpTestBase WarpPerspective;
+
+OCL_TEST_P(WarpPerspective, Mat)
+{
+    for (int j = 0; j < test_loop_times; j++)
+    {
+        random_roi();
+
+        float cols = static_cast<float>(src_roi.cols), rows = static_cast<float>(src_roi.rows);
+        float cols2 = cols / 2.0f, rows2 = rows / 2.0f;
+        Point2f sp[] = { Point2f(0.0f, 0.0f), Point2f(cols, 0.0f), Point2f(0.0f, rows), Point2f(cols, rows) };
+        Point2f dp[] = { Point2f(rng.uniform(0.0f, cols2), rng.uniform(0.0f, rows2)),
+            Point2f(rng.uniform(cols2, cols), rng.uniform(0.0f, rows2)),
+            Point2f(rng.uniform(0.0f, cols2), rng.uniform(rows2, rows)),
+            Point2f(rng.uniform(cols2, cols), rng.uniform(rows2, rows)) };
+        Mat M = getPerspectiveTransform(sp, dp);
+
+        OCL_OFF(cv::warpPerspective(src_roi, dst_roi, M, dsize, interpolation));
+        OCL_ON(cv::warpPerspective(usrc_roi, udst_roi, M, dsize, interpolation));
+
+        Near(1.0);
+    }
+}
+
+
 /////////////////////////////////////////////////////////////////////////////////////////////////
-// resize
+//// resize
 
 PARAM_TEST_CASE(Resize, MatType, double, double, Interpolation, bool)
 {
@@ -125,12 +222,118 @@ OCL_TEST_P(Resize, Mat)
     }
 }
 
+/////////////////////////////////////////////////////////////////////////////////////////////////
+// remap
+
+PARAM_TEST_CASE(Remap, MatDepth, Channels, std::pair<MatType, MatType>, Border, bool)
+{
+    int srcType, map1Type, map2Type;
+    int borderType;
+    bool useRoi;
+
+    Scalar val;
+
+    TEST_DECLARE_INPUT_PARAMETER(src)
+    TEST_DECLARE_INPUT_PARAMETER(map1)
+    TEST_DECLARE_INPUT_PARAMETER(map2)
+    TEST_DECLARE_OUTPUT_PARAMETER(dst)
+
+    virtual void SetUp()
+    {
+        srcType = CV_MAKE_TYPE(GET_PARAM(0), GET_PARAM(1));
+        map1Type = GET_PARAM(2).first;
+        map2Type = GET_PARAM(2).second;
+        borderType = GET_PARAM(3);
+        useRoi = GET_PARAM(4);
+    }
+
+    void random_roi()
+    {
+        val = randomScalar(-MAX_VALUE, MAX_VALUE);
+        Size srcROISize = randomSize(1, MAX_VALUE);
+        Size dstROISize = randomSize(1, MAX_VALUE);
+
+        Border srcBorder = randomBorder(0, useRoi ? MAX_VALUE : 0);
+        randomSubMat(src, src_roi, srcROISize, srcBorder, srcType, 5, 256);
+
+        Border dstBorder = randomBorder(0, useRoi ? MAX_VALUE : 0);
+        randomSubMat(dst, dst_roi, dstROISize, dstBorder, srcType, -MAX_VALUE, MAX_VALUE);
+
+        int mapMaxValue = MAX_VALUE << 2;
+        Border map1Border = randomBorder(0, useRoi ? MAX_VALUE : 0);
+        randomSubMat(map1, map1_roi, dstROISize, map1Border, map1Type, -mapMaxValue, mapMaxValue);
+
+        Border map2Border = randomBorder(0, useRoi ? MAX_VALUE : 0);
+        if (map2Type != noType)
+        {
+            int mapMinValue = -mapMaxValue;
+            if (map2Type == CV_16UC1 || map2Type == CV_16SC1)
+                mapMinValue = 0, mapMaxValue = INTER_TAB_SIZE2;
+            randomSubMat(map2, map2_roi, dstROISize, map2Border, map2Type, mapMinValue, mapMaxValue);
+        }
+
+        UMAT_UPLOAD_INPUT_PARAMETER(src)
+        UMAT_UPLOAD_INPUT_PARAMETER(map1)
+        UMAT_UPLOAD_OUTPUT_PARAMETER(dst)
+        if (noType != map2Type)
+            UMAT_UPLOAD_INPUT_PARAMETER(map2)
+    }
+
+    void Near(double threshold = 0.0)
+    {
+        EXPECT_MAT_NEAR(dst, udst, threshold);
+        EXPECT_MAT_NEAR(dst_roi, udst_roi, threshold);
+    }
+};
+
+typedef Remap Remap_INTER_NEAREST;
+
+OCL_TEST_P(Remap_INTER_NEAREST, Mat)
+{
+    for (int j = 0; j < test_loop_times; j++)
+    {
+        random_roi();
+
+        OCL_OFF(cv::remap(src_roi, dst_roi, map1_roi, map2_roi, INTER_NEAREST, borderType, val));
+        OCL_ON(cv::remap(usrc_roi, udst_roi, umap1_roi, umap2_roi, INTER_NEAREST, borderType, val));
+
+        Near(1.0);
+    }
+}
+
+typedef Remap Remap_INTER_LINEAR;
+
+OCL_TEST_P(Remap_INTER_LINEAR, Mat)
+{
+    for (int j = 0; j < test_loop_times; j++)
+    {
+        random_roi();
+
+        OCL_OFF(cv::remap(src_roi, dst_roi, map1_roi, map2_roi, INTER_LINEAR, borderType, val));
+        OCL_ON(cv::remap(usrc_roi, udst_roi, umap1_roi, umap2_roi, INTER_LINEAR, borderType, val));
+
+        Near(2.0);
+    }
+}
+
 /////////////////////////////////////////////////////////////////////////////////////
 
-OCL_INSTANTIATE_TEST_CASE_P(ImgprocWarpResize, Resize, Combine(
-                            Values((MatType)CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4),
-                            Values(0.7, 0.4, 2.0),
-                            Values(0.3, 0.6, 2.0),
+OCL_INSTANTIATE_TEST_CASE_P(ImgprocWarp, WarpAffine, Combine(
+                            Values(CV_8UC1, CV_8UC3, CV_8UC4, CV_32FC1, CV_32FC3, CV_32FC4),
+                            Values((Interpolation)INTER_NEAREST, (Interpolation)INTER_LINEAR, (Interpolation)INTER_CUBIC),
+                            Bool(),
+                            Bool()));
+
+OCL_INSTANTIATE_TEST_CASE_P(ImgprocWarp, WarpPerspective, Combine(
+                            Values(CV_8UC1, CV_8UC3, CV_8UC4, CV_32FC1, CV_32FC3, CV_32FC4),
+                            Values((Interpolation)INTER_NEAREST, (Interpolation)INTER_LINEAR, (Interpolation)INTER_CUBIC),
+                            Bool(),
+                            Bool()));
+
+OCL_INSTANTIATE_TEST_CASE_P(ImgprocWarp, Resize, Combine(
+                            Values(CV_8UC1, CV_8UC4, CV_16UC2, CV_32FC1, CV_32FC4),
+                            Values(0.5, 1.5, 2.0),
+                            Values(0.5, 1.5, 2.0),
                             Values((Interpolation)INTER_NEAREST, (Interpolation)INTER_LINEAR),
                             Bool()));
 
@@ -141,6 +344,33 @@ OCL_INSTANTIATE_TEST_CASE_P(ImgprocWarpResizeArea, Resize, Combine(
                             Values((Interpolation)INTER_AREA),
                             Bool()));
 
+OCL_INSTANTIATE_TEST_CASE_P(ImgprocWarp, Remap_INTER_LINEAR, Combine(
+                            Values(CV_8U, CV_16U, CV_32F),
+                            Values(1, 4),
+                            Values(std::pair<MatType, MatType>((MatType)CV_32FC1, (MatType)CV_32FC1),
+                                   std::pair<MatType, MatType>((MatType)CV_16SC2, (MatType)CV_16UC1),
+                                   std::pair<MatType, MatType>((MatType)CV_32FC2, noType)),
+                            Values((Border)BORDER_CONSTANT,
+                                   (Border)BORDER_REPLICATE,
+                                   (Border)BORDER_WRAP,
+                                   (Border)BORDER_REFLECT,
+                                   (Border)BORDER_REFLECT_101),
+                            Bool()));
+
+OCL_INSTANTIATE_TEST_CASE_P(ImgprocWarp, Remap_INTER_NEAREST, Combine(
+                            Values(CV_8U, CV_16U, CV_32F),
+                            Values(1, 4),
+                            Values(std::pair<MatType, MatType>((MatType)CV_32FC1, (MatType)CV_32FC1),
+                                   std::pair<MatType, MatType>((MatType)CV_32FC2, noType),
+                                   std::pair<MatType, MatType>((MatType)CV_16SC2, (MatType)CV_16UC1),
+                                   std::pair<MatType, MatType>((MatType)CV_16SC2, noType)),
+                            Values((Border)BORDER_CONSTANT,
+                                   (Border)BORDER_REPLICATE,
+                                   (Border)BORDER_WRAP,
+                                   (Border)BORDER_REFLECT,
+                                   (Border)BORDER_REFLECT_101),
+                            Bool()));
+
 } } // namespace cvtest::ocl
 
 #endif // HAVE_OPENCL

modules/ts/include/opencv2/ts/ocl_test.hpp

@@ -306,6 +306,7 @@ IMPLEMENT_PARAM_CLASS(Channels, int)
 #define OCL_ALL_CHANNELS Values(1, 2, 3, 4)
 
 CV_ENUM(Interpolation, INTER_NEAREST, INTER_LINEAR, INTER_CUBIC, INTER_AREA)
+CV_ENUM(Border, BORDER_CONSTANT, BORDER_REPLICATE, BORDER_WRAP, BORDER_REFLECT, BORDER_REFLECT_101)
 
 #define OCL_INSTANTIATE_TEST_CASE_P(prefix, test_case_name, generator) \
     INSTANTIATE_TEST_CASE_P(OCL_ ## prefix, test_case_name, generator)