Merge pull request #2897 from vbystricky:oclopt_sepFilter2D

modules/imgproc/src/filter.cpp

@@ -3471,7 +3471,8 @@ static bool ocl_sepColFilter2D(const UMat & buf, UMat & dst, const Mat & kernelY
     return k.run(2, globalsize, localsize, false);
 }
 
-const int optimizedSepFilterLocalSize = 16;
+const int optimizedSepFilterLocalWidth  = 16;
+const int optimizedSepFilterLocalHeight = 8;
 
 static bool ocl_sepFilter2D_SinglePass(InputArray _src, OutputArray _dst,
                                        Mat row_kernel, Mat col_kernel,
@@ -3491,8 +3492,8 @@ static bool ocl_sepFilter2D_SinglePass(InputArray _src, OutputArray _dst,
               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]) };
+    size_t lt2[2] = { optimizedSepFilterLocalWidth, optimizedSepFilterLocalHeight };
+    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",
@@ -3584,8 +3585,8 @@ static bool ocl_sepFilter2D( InputArray _src, OutputArray _dst, int ddepth,
     }
 
     CV_OCL_RUN_(kernelY.cols <= 21 && kernelX.cols <= 21 &&
-                imgSize.width > optimizedSepFilterLocalSize + anchor.x &&
-                imgSize.height > optimizedSepFilterLocalSize + anchor.y &&
+                imgSize.width > optimizedSepFilterLocalWidth + anchor.x &&
+                imgSize.height > optimizedSepFilterLocalHeight + anchor.y &&
                 (!(borderType & BORDER_ISOLATED) || _src.offset() == 0) &&
                 anchor == Point(kernelX.cols >> 1, kernelY.cols >> 1) &&
                 (d.isIntel() || (d.isAMD() && !d.hostUnifiedMemory())),

modules/imgproc/src/opencl/filterSep_singlePass.cl

@@ -119,20 +119,17 @@ __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int
     int liy = get_local_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;
-    int srcY = y + srcOffsetY - RADIUSY;
 
     // 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;
+        int yb = clocY + srcOffsetY - RADIUSY;
         EXTRAPOLATE(yb, (height));
 
         int clocX = lix;
@@ -149,11 +146,12 @@ __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int
         while(clocX < BLK_X+(RADIUSX*2));
 
         clocY += BLK_Y;
-        cSrcY += BLK_Y;
     }
     while (clocY < BLK_Y+(RADIUSY*2));
     barrier(CLK_LOCAL_MEM_FENCE);
 
+    for (int y = 0; y < dst_rows; y+=BLK_Y)
+    {
         // do vertical filter pass
         // and store intermediate results to second local memory array
         int i, clocX = lix;
@@ -163,9 +161,9 @@ __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int
             sum = (WT) 0;
             for (i=0; i<=2*RADIUSY; i++)
 #if (defined(INTEGER_ARITHMETIC) && !INTEL_DEVICE)
-            sum = mad24(lsmem[liy+i][clocX], mat_kernelY[i], sum);
+                sum = mad24(lsmem[liy + i][clocX], mat_kernelY[i], sum);
 #else
-            sum = mad(lsmem[liy+i][clocX], mat_kernelY[i], sum);
+                sum = mad(lsmem[liy + i][clocX], mat_kernelY[i], sum);
 #endif
             lsmemDy[liy][clocX] = sum;
             clocX += BLK_X;
@@ -175,9 +173,8 @@ __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int
 
         // 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;
-
+        if ((x < dst_cols) && (y + liy < dst_rows))
+        {
             // do second horizontal filter pass
             // and calculate final result
             sum = 0.0f;
@@ -195,7 +192,34 @@ __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int
             sum = (sum + (1 << (SHIFT_BITS-1))) >> SHIFT_BITS;
 #endif
 #endif
-
             // store result into destination image
-    storepix(convertToDstT(sum + (WT)(delta)), Dst + mad24(y, dst_step, mad24(x, DSTSIZE, dst_offset)));
+            storepix(convertToDstT(sum + (WT)(delta)), Dst + mad24(y + liy, dst_step, mad24(x, DSTSIZE, dst_offset)));
+        }
+
+        for (int i = liy * BLK_X + lix; i < (RADIUSY*2) * (BLK_X+(RADIUSX*2)); i += BLK_X * BLK_Y)
+        {
+            int clocX = i % (BLK_X+(RADIUSX*2));
+            int clocY = i / (BLK_X+(RADIUSX*2));
+            lsmem[clocY][clocX] = lsmem[clocY + BLK_Y][clocX];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        int yb = y + liy + BLK_Y + srcOffsetY + RADIUSY;
+        EXTRAPOLATE(yb, (height));
+
+        clocX = lix;
+        int cSrcX = x + srcOffsetX - RADIUSX;
+        do
+        {
+            int xb = cSrcX;
+            EXTRAPOLATE(xb,(width));
+            lsmem[liy + 2*RADIUSY][clocX] = ELEM(xb, yb, (width), (height), 0 );
+
+            clocX += BLK_X;
+            cSrcX += BLK_X;
+        }
+        while(clocX < BLK_X+(RADIUSX*2));
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
 }