simplified cv::sepFilter2D OpenCL part

modules/imgproc/src/opencl/filterSepCol.cl

@@ -36,16 +36,6 @@
 
 #define READ_TIMES_COL ((2*(RADIUSY+LSIZE1)-1)/LSIZE1)
 #define RADIUS 1
-#if CN ==1
-#define ALIGN (((RADIUS)+3)>>2<<2)
-#elif CN==2
-#define ALIGN (((RADIUS)+1)>>1<<1)
-#elif CN==3
-#define ALIGN (((RADIUS)+3)>>2<<2)
-#elif CN==4
-#define ALIGN (RADIUS)
-#define READ_TIMES_ROW ((2*(RADIUS+LSIZE0)-1)/LSIZE0)
-#endif
 
 #define noconvert
 
@@ -65,16 +55,8 @@ The info above maybe obsolete.
 #define DIG(a) a,
 __constant float mat_kernel[] = { COEFF };
 
-__kernel __attribute__((reqd_work_group_size(LSIZE0,LSIZE1,1))) void col_filter
-                        (__global const GENTYPE_SRC * restrict src,
-                         const int src_step_in_pixel,
-                         const int src_whole_cols,
-                         const int src_whole_rows,
-                         __global GENTYPE_DST * dst,
-                         const int dst_offset_in_pixel,
-                         const int dst_step_in_pixel,
-                         const int dst_cols,
-                         const int dst_rows)
+__kernel void col_filter(__global const srcT * src, int src_step_in_pixel, int src_whole_cols, int src_whole_rows,
+                         __global dstT * dst, int dst_offset_in_pixel, int dst_step_in_pixel, int dst_cols, int dst_rows)
 {
     int x = get_global_id(0);
     int y = get_global_id(1);
@@ -85,35 +67,35 @@ __kernel __attribute__((reqd_work_group_size(LSIZE0,LSIZE1,1))) void col_filter
     int start_addr = mad24(y, src_step_in_pixel, x);
     int end_addr = mad24(src_whole_rows - 1, src_step_in_pixel, src_whole_cols);
 
-    int i;
-    GENTYPE_SRC sum, temp[READ_TIMES_COL];
-    __local GENTYPE_SRC LDS_DAT[LSIZE1 * READ_TIMES_COL][LSIZE0 + 1];
+    srcT sum, temp[READ_TIMES_COL];
+    __local srcT LDS_DAT[LSIZE1 * READ_TIMES_COL][LSIZE0 + 1];
 
-    //read pixels from src
-    for(i = 0;i<READ_TIMES_COL;i++)
+    // read pixels from src
+    for (int i = 0; i < READ_TIMES_COL; ++i)
     {
-        int current_addr = start_addr+i*LSIZE1*src_step_in_pixel;
+        int current_addr = mad24(i, LSIZE1 * src_step_in_pixel, start_addr);
         current_addr = current_addr < end_addr ? current_addr : 0;
         temp[i] = src[current_addr];
     }
-    //save pixels to lds
-    for(i = 0;i<READ_TIMES_COL;i++)
-    {
-        LDS_DAT[l_y+i*LSIZE1][l_x] = temp[i];
-    }
+
+    // save pixels to lds
+    for (int i = 0; i < READ_TIMES_COL; ++i)
+        LDS_DAT[mad24(i, LSIZE1, l_y)][l_x] = temp[i];
     barrier(CLK_LOCAL_MEM_FENCE);
-    //read pixels from lds and calculate the result
-    sum = LDS_DAT[l_y+RADIUSY][l_x]*mat_kernel[RADIUSY];
-    for(i=1;i<=RADIUSY;i++)
+
+    // read pixels from lds and calculate the result
+    sum = LDS_DAT[l_y + RADIUSY][l_x] * mat_kernel[RADIUSY];
+    for (int i = 1; i <= RADIUSY; ++i)
     {
-        temp[0]=LDS_DAT[l_y+RADIUSY-i][l_x];
-        temp[1]=LDS_DAT[l_y+RADIUSY+i][l_x];
-        sum += temp[0] * mat_kernel[RADIUSY-i]+temp[1] * mat_kernel[RADIUSY+i];
+        temp[0] = LDS_DAT[l_y + RADIUSY - i][l_x];
+        temp[1] = LDS_DAT[l_y + RADIUSY + i][l_x];
+        sum += mad(temp[0], mat_kernel[RADIUSY - i], temp[1] * mat_kernel[RADIUSY + i]);
     }
-    //write the result to dst
-    if((x<dst_cols) & (y<dst_rows))
+
+    // write the result to dst
+    if (x < dst_cols && y < dst_rows)
     {
         start_addr = mad24(y, dst_step_in_pixel, x + dst_offset_in_pixel);
-        dst[start_addr] = convert_to_DST(sum);
+        dst[start_addr] = convertToDstT(sum);
     }
 }

modules/imgproc/src/opencl/filterSep_singlePass.cl

@@ -75,6 +75,7 @@
 #endif
 
 #define SRC(_x,_y) convertToWT(((global srcT*)(Src+(_y)*src_step))[_x])
+#define DST(_x,_y) (((global dstT*)(Dst+dst_offset+(_y)*dst_step))[_x])
 
 #ifdef BORDER_CONSTANT
 // CCCCCC|abcdefgh|CCCCCCC
@@ -83,8 +84,6 @@
 #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
@@ -101,15 +100,15 @@ __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int
     // 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];
+    __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);
+    int x = get_global_id(0);
+    int y = get_global_id(1);
 
     // calculate pixel position in source image taking image offset into account
     int srcX = x + srcOffsetX - RADIUSX;

modules/imgproc/test/ocl/test_sepfilter2D.cpp

@@ -79,12 +79,14 @@ PARAM_TEST_CASE(SepFilter2D, MatDepth, Channels, BorderType, bool, bool)
             ksize.width++;
         if (1 != (ksize.height % 2))
             ksize.height++;
+
         Mat temp = randomMat(Size(ksize.width, 1), CV_MAKE_TYPE(CV_32F, 1), -MAX_VALUE, MAX_VALUE);
         cv::normalize(temp, kernelX, 1.0, 0.0, NORM_L1);
         temp = randomMat(Size(1, ksize.height),  CV_MAKE_TYPE(CV_32F, 1), -MAX_VALUE, MAX_VALUE);
         cv::normalize(temp, kernelY, 1.0, 0.0, NORM_L1);
 
-        Size roiSize = randomSize(ksize.width, MAX_VALUE, ksize.height, MAX_VALUE);
+        Size roiSize = randomSize(ksize.width + 16, MAX_VALUE, ksize.height + 20, MAX_VALUE);
+        std::cout << roiSize << std::endl;
         int rest = roiSize.width % 4;
         if (0 != rest)
             roiSize.width += (4 - rest);