optimized sep filter

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

@@ -592,7 +592,7 @@ protected:
 CV_EXPORTS const char* convertTypeStr(int sdepth, int ddepth, int cn, char* buf);
 CV_EXPORTS const char* typeToStr(int t);
 CV_EXPORTS const char* memopTypeToStr(int t);
-CV_EXPORTS String kernelToStr(InputArray _kernel, int ddepth = -1);
+CV_EXPORTS String kernelToStr(InputArray _kernel, int ddepth = -1, const char * name = NULL);
 CV_EXPORTS void getPlatfomsInfo(std::vector<PlatformInfo>& platform_info);
 CV_EXPORTS int predictOptimalVectorWidth(InputArray src1, InputArray src2 = noArray(), InputArray src3 = noArray(),
                                          InputArray src4 = noArray(), InputArray src5 = noArray(), InputArray src6 = noArray(),

modules/core/src/ocl.cpp

@@ -4306,7 +4306,7 @@ static std::string kerToStr(const Mat & k)
     return stream.str();
 }
 
-String kernelToStr(InputArray _kernel, int ddepth)
+String kernelToStr(InputArray _kernel, int ddepth, const char * name)
 {
     Mat kernel = _kernel.getMat().reshape(1, 1);
 
@@ -4323,7 +4323,7 @@ String kernelToStr(InputArray _kernel, int ddepth)
     const func_t func = funcs[depth];
     CV_Assert(func != 0);
 
-    return cv::format(" -D COEFF=%s", func(kernel).c_str());
+    return cv::format(" -D %s=%s", name ? name : "COEFF", func(kernel).c_str());
 }
 
 #define PROCESS_SRC(src) \

modules/imgproc/perf/opencl/perf_filters.cpp

@@ -211,7 +211,7 @@ OCL_PERF_TEST_P(SobelFixture, Sobel,
 
     OCL_TEST_CYCLE() cv::Sobel(src, dst, -1, dx, dy);
 
-    SANITY_CHECK(dst);
+    SANITY_CHECK(dst, 1e-6);
 }
 
 ///////////// Scharr ////////////////////////

modules/imgproc/src/filter.cpp

@@ -3350,27 +3350,8 @@ static bool ocl_sepRowFilter2D( UMat &src, UMat &buf, Mat &kernelX, int anchor,
     int radiusY = (int)((buf.rows - src.rows) >> 1);
 
     bool isIsolatedBorder = (borderType & BORDER_ISOLATED) != 0;
-    const char* btype = NULL;
-    switch (borderType & ~BORDER_ISOLATED)
-    {
-    case BORDER_CONSTANT:
-        btype = "BORDER_CONSTANT";
-        break;
-    case BORDER_REPLICATE:
-        btype = "BORDER_REPLICATE";
-        break;
-    case BORDER_REFLECT:
-        btype = "BORDER_REFLECT";
-        break;
-    case BORDER_WRAP:
-        btype = "BORDER_WRAP";
-        break;
-    case BORDER_REFLECT101:
-        btype = "BORDER_REFLECT_101";
-        break;
-    default:
-        return false;
-    }
+    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", "BORDER_WRAP", "BORDER_REFLECT_101" },
+        * const btype = borderMap[borderType & ~BORDER_ISOLATED];
 
     bool extra_extrapolation = src.rows < (int)((-radiusY + globalsize[1]) >> 1) + 1;
     extra_extrapolation |= src.rows < radiusY;
@@ -3463,36 +3444,96 @@ static bool ocl_sepColFilter2D(const UMat &buf, UMat &dst, Mat &kernelY, int anc
     return kernelCol.run(2, globalsize, localsize, sync);
 }
 
+const int optimizedSepFilterLocalSize = 16;
+
+static bool ocl_sepFilter2D_SinglePass(InputArray _src, OutputArray _dst,
+                                       InputArray _row_kernel, InputArray _col_kernel,
+                                       int borderType, int ddepth)
+{
+    Size size = _src.size(), wholeSize;
+    Point origin;
+    int stype = _src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype),
+            esz = CV_ELEM_SIZE(stype), wdepth = std::max(std::max(sdepth, ddepth), CV_32F),
+            dtype = CV_MAKE_TYPE(ddepth, cn);
+    size_t src_step = _src.step(), src_offset = _src.offset();
+    bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
+
+    if ((src_offset % src_step) % esz != 0 || (!doubleSupport && sdepth == CV_64F) ||
+            !(borderType == BORDER_CONSTANT || borderType == BORDER_REPLICATE ||
+              borderType == BORDER_REFLECT || borderType == BORDER_WRAP ||
+              borderType == BORDER_REFLECT_101))
+        return false;
+
+    size_t lt2[2] = { optimizedSepFilterLocalSize, optimizedSepFilterLocalSize };
+    size_t gt2[2] = { lt2[0] * (1 + (size.width - 1) / lt2[0]), lt2[1] * (1 + (size.height - 1) / lt2[1]) };
+
+    char cvt[2][40];
+    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", "BORDER_WRAP",
+                                       "BORDER_REFLECT_101" };
+
+    String opts = cv::format("-D BLK_X=%d -D BLK_Y=%d -D RADIUSX=%d -D RADIUSY=%d%s%s"
+                             " -D srcT=%s -D convertToWT=%s -D WT=%s -D dstT=%s -D convertToDstT=%s"
+                             " -D %s", (int)lt2[0], (int)lt2[1], _row_kernel.size().height / 2, _col_kernel.size().height / 2,
+                             ocl::kernelToStr(_row_kernel, CV_32F, "KERNEL_MATRIX_X").c_str(),
+                             ocl::kernelToStr(_col_kernel, CV_32F, "KERNEL_MATRIX_Y").c_str(),
+                             ocl::typeToStr(stype), ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]),
+                             ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)), ocl::typeToStr(dtype),
+                             ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]), borderMap[borderType]);
+
+    ocl::Kernel k("sep_filter", ocl::imgproc::filterSep_singlePass_oclsrc, opts);
+    if (k.empty())
+        return false;
+
+    UMat src = _src.getUMat();
+    _dst.create(size, dtype);
+    UMat dst = _dst.getUMat();
+
+    int src_offset_x = static_cast<int>((src_offset % src_step) / esz);
+    int src_offset_y = static_cast<int>(src_offset / src_step);
+
+    src.locateROI(wholeSize, origin);
+
+    k.args(ocl::KernelArg::PtrReadOnly(src), (int)src_step, src_offset_x, src_offset_y,
+           wholeSize.height, wholeSize.width, ocl::KernelArg::WriteOnly(dst));
+
+    return k.run(2, gt2, lt2, false);
+}
+
 static bool ocl_sepFilter2D( InputArray _src, OutputArray _dst, int ddepth,
                       InputArray _kernelX, InputArray _kernelY, Point anchor,
                       double delta, int borderType )
 {
+    Size imgSize = _src.size();
+
     if (abs(delta)> FLT_MIN)
         return false;
 
-    int type = _src.type();
+    int type = _src.type(), cn = CV_MAT_CN(type);
     if ( !( (type == CV_8UC1 || type == CV_8UC4 || type == CV_32FC1 || type == CV_32FC4) &&
             (ddepth == CV_32F || ddepth == CV_16S || ddepth == CV_8U || ddepth < 0) ) )
         return false;
 
-    int cn = CV_MAT_CN(type);
-
     Mat kernelX = _kernelX.getMat().reshape(1, 1);
-    if (1 != (kernelX.cols % 2))
+    if (kernelX.cols % 2 != 1)
         return false;
     Mat kernelY = _kernelY.getMat().reshape(1, 1);
-    if (1 != (kernelY.cols % 2))
+    if (kernelY.cols % 2 != 1)
         return false;
 
     int sdepth = CV_MAT_DEPTH(type);
-    if( anchor.x < 0 )
+    if (anchor.x < 0)
         anchor.x = kernelX.cols >> 1;
-    if( anchor.y < 0 )
+    if (anchor.y < 0)
         anchor.y = kernelY.cols >> 1;
 
-    if( ddepth < 0 )
+    if (ddepth < 0)
         ddepth = sdepth;
 
+    CV_OCL_RUN_(kernelY.rows <= 21 && kernelX.rows <= 21 &&
+        imgSize.width > optimizedSepFilterLocalSize + (kernelX.rows >> 1) &&
+        imgSize.height > optimizedSepFilterLocalSize + (kernelY.rows >> 1),
+        ocl_sepFilter2D_SinglePass(_src, _dst, _kernelX, _kernelY, borderType, ddepth), true)
+
     UMat src = _src.getUMat();
     Size srcWholeSize; Point srcOffset;
     src.locateROI(srcWholeSize, srcOffset);

modules/imgproc/src/opencl/filterSep_singlePass.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) 2014, Intel Corporation, 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*/
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+/////////////////////////////////Macro for border type////////////////////////////////////////////
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+#ifdef BORDER_CONSTANT
+// CCCCCC|abcdefgh|CCCCCCC
+#define EXTRAPOLATE(x, maxV)
+#elif defined BORDER_REPLICATE
+// aaaaaa|abcdefgh|hhhhhhh
+#define EXTRAPOLATE(x, maxV) \
+    { \
+        (x) = max(min((x), (maxV) - 1), 0); \
+    }
+#elif defined BORDER_WRAP
+// cdefgh|abcdefgh|abcdefg
+#define EXTRAPOLATE(x, maxV) \
+    { \
+        (x) = ( (x) + (maxV) ) % (maxV); \
+    }
+#elif defined BORDER_REFLECT
+// fedcba|abcdefgh|hgfedcb
+#define EXTRAPOLATE(x, maxV) \
+    { \
+        (x) = min(((maxV)-1)*2-(x)+1, max((x),-(x)-1) ); \
+    }
+#elif defined BORDER_REFLECT_101 || defined BORDER_REFLECT101
+// gfedcb|abcdefgh|gfedcba
+#define EXTRAPOLATE(x, maxV) \
+    { \
+        (x) = min(((maxV)-1)*2-(x), max((x),-(x)) ); \
+    }
+#else
+#error No extrapolation method
+#endif
+
+#define SRC(_x,_y) convertToWT(((global srcT*)(Src+(_y)*src_step))[_x])
+
+#ifdef BORDER_CONSTANT
+// CCCCCC|abcdefgh|CCCCCCC
+#define ELEM(_x,_y,r_edge,t_edge,const_v) (_x)<0 | (_x) >= (r_edge) | (_y)<0 | (_y) >= (t_edge) ? (const_v) : SRC((_x),(_y))
+#else
+#define ELEM(_x,_y,r_edge,t_edge,const_v) SRC((_x),(_y))
+#endif
+
+#define DST(_x,_y) (((global dstT*)(Dst+dst_offset+(_y)*dst_step))[_x])
+
+#define noconvert
+
+// horizontal and vertical filter kernels
+// should be defined on host during compile time to avoid overhead
+#define DIG(a) a,
+__constant float mat_kernelX[] = { KERNEL_MATRIX_X };
+__constant float mat_kernelY[] = { KERNEL_MATRIX_Y };
+
+__kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int srcOffsetY, int height, int width,
+                         __global uchar* Dst, int dst_step, int dst_offset, int dst_rows, int dst_cols)
+{
+    // RADIUSX, RADIUSY are filter dimensions
+    // BLK_X, BLK_Y are local wrogroup sizes
+    // all these should be defined on host during compile time
+    // first lsmem array for source pixels used in first pass,
+    // second lsmemDy for storing first pass results
+    __local WT lsmem[BLK_Y+2*RADIUSY][BLK_X+2*RADIUSX];
+    __local WT lsmemDy[BLK_Y][BLK_X+2*RADIUSX];
+
+    // get local and global ids - used as image and local memory array indexes
+    int lix = get_local_id(0);
+    int liy = get_local_id(1);
+
+    int x = (int)get_global_id(0);
+    int y = (int)get_global_id(1);
+
+    // calculate pixel position in source image taking image offset into account
+    int srcX = x + srcOffsetX - RADIUSX;
+    int srcY = y + srcOffsetY - RADIUSY;
+    int xb = srcX;
+    int yb = srcY;
+
+    // extrapolate coordinates, if needed
+    // and read my own source pixel into local memory
+    // with account for extra border pixels, which will be read by starting workitems
+    int clocY = liy;
+    int cSrcY = srcY;
+    do
+    {
+        int yb = cSrcY;
+        EXTRAPOLATE(yb, (height));
+
+        int clocX = lix;
+        int cSrcX = srcX;
+        do
+        {
+            int xb = cSrcX;
+            EXTRAPOLATE(xb,(width));
+            lsmem[clocY][clocX] = ELEM(xb, yb, (width), (height), 0 );
+
+            clocX += BLK_X;
+            cSrcX += BLK_X;
+        }
+        while(clocX < BLK_X+(RADIUSX*2));
+
+        clocY += BLK_Y;
+        cSrcY += BLK_Y;
+    }
+    while (clocY < BLK_Y+(RADIUSY*2));
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    // do vertical filter pass
+    // and store intermediate results to second local memory array
+    int i, clocX = lix;
+    WT sum = 0.0f;
+    do
+    {
+        sum = 0.0f;
+        for (i=0; i<=2*RADIUSY; i++)
+            sum = mad(lsmem[liy+i][clocX], mat_kernelY[i], sum);
+        lsmemDy[liy][clocX] = sum;
+        clocX += BLK_X;
+    }
+    while(clocX < BLK_X+(RADIUSX*2));
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    // if this pixel happened to be out of image borders because of global size rounding,
+    // then just return
+    if( x >= dst_cols || y >=dst_rows )
+        return;
+
+    // do second horizontal filter pass
+    // and calculate final result
+    sum = 0.0f;
+    for (i=0; i<=2*RADIUSX; i++)
+        sum = mad(lsmemDy[liy][lix+i], mat_kernelX[i], sum);
+
+    //store result into destination image
+    DST(x,y) = convertToDstT(sum);
+}