Optimize OpenCL version of Laplacian filter for kernel size great than 3

modules/imgproc/src/deriv.cpp

@@ -643,21 +643,103 @@ void cv::Scharr( InputArray _src, OutputArray _dst, int ddepth, int dx, int dy,
 
 namespace cv {
 
+#define LAPLACIAN_LOCAL_MEM(tileX, tileY, ksize, elsize) (((tileX) + 2 * (int)((ksize) / 2)) * (3 * (tileY) + 2 * (int)((ksize) / 2)) * elsize)
+
 static bool ocl_Laplacian5(InputArray _src, OutputArray _dst,
                            const Mat & kd, const Mat & ks, double scale, double delta,
                            int borderType, int depth, int ddepth)
 {
+    const size_t tileSizeX = 16;
+    const size_t tileSizeYmin = 8;
+
+    const ocl::Device dev = ocl::Device::getDefault();
+
+    int stype = _src.type();
+    int sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype), esz = CV_ELEM_SIZE(stype);
+
+    bool doubleSupport = dev.doubleFPConfig() > 0;
+    if (!doubleSupport && (sdepth == CV_64F || ddepth == CV_64F))
+        return false;
+
+    Mat kernelX = kd.reshape(1, 1);
+    if (kernelX.cols % 2 != 1)
+        return false;
+    Mat kernelY = ks.reshape(1, 1);
+    if (kernelY.cols % 2 != 1)
+        return false;
+    CV_Assert(kernelX.cols == kernelY.cols);
+
+    size_t wgs = dev.maxWorkGroupSize();
+    size_t lmsz = dev.localMemSize();
+
+    size_t src_step = _src.step(), src_offset = _src.offset();
+    if (((src_offset % src_step) % esz == 0) &&
+        (
+         (borderType == BORDER_CONSTANT || borderType == BORDER_REPLICATE) ||
+         (borderType == BORDER_REFLECT || borderType == BORDER_WRAP || borderType == BORDER_REFLECT_101) &&
+         (_src.cols() >= kernelX.cols && _src.rows() >= kernelY.cols)
+        ) &&
+        (tileSizeX * tileSizeYmin  <= wgs) &&
+        (LAPLACIAN_LOCAL_MEM(tileSizeX, tileSizeYmin, kernelX.cols, cn * 4) <= lmsz)
+       )
+    {
+        Size size = _src.size(), wholeSize;
+        Point origin;
+        int dtype = CV_MAKE_TYPE(ddepth, cn);
+        int wdepth = CV_32F;
+
+        size_t tileSizeY = wgs / tileSizeX;
+        while ((tileSizeX * tileSizeY > wgs) || (LAPLACIAN_LOCAL_MEM(tileSizeX, tileSizeY, kernelX.cols, cn * 4) > lmsz))
+        {
+            tileSizeY /= 2;
+        }
+        size_t lt2[2] = { tileSizeX, tileSizeY};
+        size_t gt2[2] = { lt2[0] * (1 + (size.width - 1) / lt2[0]), 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 RADIUS=%d%s%s"
+                                 " -D convertToWT=%s -D convertToDT=%s"
+                                 " -D %s -D srcT1=%s -D dstT1=%s -D WT1=%s -D CN=%d ",
+                                 (int)lt2[0], (int)lt2[1], kernelX.cols / 2,
+                                 ocl::kernelToStr(kernelX, wdepth, "KERNEL_MATRIX_X").c_str(),
+                                 ocl::kernelToStr(kernelY, wdepth, "KERNEL_MATRIX_Y").c_str(),
+                                 ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]),
+                                 ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]),
+                                 borderMap[borderType], ocl::typeToStr(sdepth), ocl::typeToStr(ddepth), ocl::typeToStr(wdepth),
+                                 cn);
+
+        ocl::Kernel k("laplacian", ocl::imgproc::laplacian5_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),
+               static_cast<float>(scale), static_cast<float>(delta));
+
+        return k.run(2, gt2, lt2, false);
+    }
     int iscale = cvRound(scale), idelta = cvRound(delta);
-    bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0,
-            floatCoeff = std::fabs(delta - idelta) > DBL_EPSILON || std::fabs(scale - iscale) > DBL_EPSILON;
-    int cn = _src.channels(), wdepth = std::max(depth, floatCoeff ? CV_32F : CV_32S), kercn = 1;
+    bool floatCoeff = std::fabs(delta - idelta) > DBL_EPSILON || std::fabs(scale - iscale) > DBL_EPSILON;
+    int wdepth = std::max(depth, floatCoeff ? CV_32F : CV_32S), kercn = 1;
 
     if (!doubleSupport && wdepth == CV_64F)
         return false;
 
     char cvt[2][40];
     ocl::Kernel k("sumConvert", ocl::imgproc::laplacian5_oclsrc,
-                  format("-D srcT=%s -D WT=%s -D dstT=%s -D coeffT=%s -D wdepth=%d "
+                  format("-D ONLY_SUM_CONVERT "
+                         "-D srcT=%s -D WT=%s -D dstT=%s -D coeffT=%s -D wdepth=%d "
                          "-D convertToWT=%s -D convertToDT=%s%s",
                          ocl::typeToStr(CV_MAKE_TYPE(depth, kercn)),
                          ocl::typeToStr(CV_MAKE_TYPE(wdepth, kercn)),

modules/imgproc/src/opencl/laplacian5.cl

@@ -5,8 +5,11 @@
 // Copyright (C) 2014, Itseez, Inc., all rights reserved.
 // Third party copyrights are property of their respective owners.
 
+
 #define noconvert
 
+#ifdef ONLY_SUM_CONVERT
+
 __kernel void sumConvert(__global const uchar * src1ptr, int src1_step, int src1_offset,
                          __global const uchar * src2ptr, int src2_step, int src2_offset,
                          __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols,
@@ -32,3 +35,185 @@ __kernel void sumConvert(__global const uchar * src1ptr, int src1_step, int src1
 #endif
     }
 }
+
+#else
+
+#if CN==1
+#define srcT    srcT1
+#define WT      WT1
+#define dstT    dstT1
+#else
+#define __CAT(x, y) x##y
+#define CAT(x, y) __CAT(x, y)
+
+#define srcT    CAT(srcT1, CN)
+#define WT      CAT(WT1, CN)
+#define dstT    CAT(dstT1, CN)
+#endif
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+/////////////////////////////////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
+// gfedcb|abcdefgh|gfedcba
+#define EXTRAPOLATE(x, maxV) \
+    { \
+        (x) = min(((maxV)-1)*2-(x), max((x),-(x)) ); \
+    }
+#else
+#error No extrapolation method
+#endif
+
+#if CN != 3
+#define loadpix(addr) *(__global const srcT *)(addr)
+#define storepix(val, addr)  *(__global dstT *)(addr) = val
+#define SRCSIZE (int)sizeof(srcT)
+#define DSTSIZE (int)sizeof(dstT)
+#else
+#define loadpix(addr)  vload3(0, (__global const srcT1 *)(addr))
+#define storepix(val, addr) vstore3(val, 0, (__global dstT1 *)(addr))
+#define SRCSIZE (int)sizeof(srcT1)*3
+#define DSTSIZE (int)sizeof(dstT1)*3
+#endif
+
+#define SRC(_x,_y) convertToWT(loadpix(Src + mad24(_y, src_step, SRCSIZE * _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
+
+// horizontal and vertical filter kernels
+// should be defined on host during compile time to avoid overhead
+#define DIG(a) a,
+__constant WT1 mat_kernelX[] = { KERNEL_MATRIX_X };
+__constant WT1 mat_kernelY[] = { KERNEL_MATRIX_Y };
+
+__kernel void laplacian(__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,
+                         WT1 scale, WT1 delta)
+{
+    __local WT lsmem[BLK_Y + 2 * RADIUS][BLK_X + 2 * RADIUS];
+    __local WT lsmemDy1[BLK_Y][BLK_X + 2 * RADIUS];
+    __local WT lsmemDy2[BLK_Y][BLK_X + 2 * RADIUS];
+
+    int lix = get_local_id(0);
+    int liy = get_local_id(1);
+
+    int x = get_global_id(0);
+
+    int srcX = x + srcOffsetX - RADIUS;
+
+    int clocY = liy;
+    do
+    {
+        int yb = clocY + srcOffsetY - RADIUS;
+        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+(RADIUS*2));
+
+        clocY += BLK_Y;
+    }
+    while (clocY < BLK_Y+(RADIUS*2));
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    WT scale_v = (WT)scale;
+    WT delta_v = (WT)delta;
+    for (int y = 0; y < dst_rows; y+=BLK_Y)
+    {
+        int i, clocX = lix;
+        WT sum1 = (WT) 0;
+        WT sum2 = (WT) 0;
+        do
+        {
+            sum1 = (WT) 0;
+            sum2 = (WT) 0;
+            for (i=0; i<=2*RADIUS; i++)
+            {
+                sum1 = mad(lsmem[liy + i][clocX], mat_kernelY[i], sum1);
+                sum2 = mad(lsmem[liy + i][clocX], mat_kernelX[i], sum2);
+            }
+            lsmemDy1[liy][clocX] = sum1;
+            lsmemDy2[liy][clocX] = sum2;
+            clocX += BLK_X;
+        }
+        while(clocX < BLK_X+(RADIUS*2));
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if ((x < dst_cols) && (y + liy < dst_rows))
+        {
+            sum1 = (WT) 0;
+            sum2 = (WT) 0;
+            for (i=0; i<=2*RADIUS; i++)
+            {
+                sum1 = mad(lsmemDy1[liy][lix+i], mat_kernelX[i], sum1);
+                sum2 = mad(lsmemDy2[liy][lix+i], mat_kernelY[i], sum2);
+            }
+
+            WT sum = mad(scale_v, (sum1 + sum2), delta_v);
+            storepix(convertToDT(sum), Dst + mad24(y + liy, dst_step, mad24(x, DSTSIZE, dst_offset)));
+        }
+
+        for (int i = liy * BLK_X + lix; i < (RADIUS*2) * (BLK_X+(RADIUS*2)); i += BLK_X * BLK_Y)
+        {
+            int clocX = i % (BLK_X+(RADIUS*2));
+            int clocY = i / (BLK_X+(RADIUS*2));
+            lsmem[clocY][clocX] = lsmem[clocY + BLK_Y][clocX];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        int yb = y + liy + BLK_Y + srcOffsetY + RADIUS;
+        EXTRAPOLATE(yb, (height));
+
+        clocX = lix;
+        int cSrcX = x + srcOffsetX - RADIUS;
+        do
+        {
+            int xb = cSrcX;
+            EXTRAPOLATE(xb,(width));
+            lsmem[liy + 2*RADIUS][clocX] = ELEM(xb, yb, (width), (height), 0 );
+
+            clocX += BLK_X;
+            cSrcX += BLK_X;
+        }
+        while(clocX < BLK_X+(RADIUS*2));
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+}
+
+#endif
\ No newline at end of file