fix bug of ocl::canny

modules/ocl/src/kernels/imgproc_canny.cl

@@ -100,8 +100,6 @@ __kernel
 
     __local int smem[16][18];
 
-    if(gidy < rows)
-    {
     smem[lidy][lidx + 1] = src[gidx + gidy * src_step + src_offset]; 
     if(lidx == 0)
     {
@@ -109,6 +107,10 @@ __kernel
         smem[lidy][17] = src[min(gidx + 16, cols - 1) + gidy * src_step + src_offset]; 
     }
     barrier(CLK_LOCAL_MEM_FENCE);
+
+    if(gidy < rows)
+    {
+
         if(gidx < cols)
         {
             dx_buf[gidx + gidy * dx_buf_step + dx_buf_offset] =
@@ -169,8 +171,6 @@ __kernel
     __local int sdx[18][16];
     __local int sdy[18][16];
     
-    if(gidx < cols)
-    {
     sdx[lidy + 1][lidx] = dx_buf[gidx + gidy * dx_buf_step + dx_buf_offset];
     sdy[lidy + 1][lidx] = dy_buf[gidx + gidy * dy_buf_step + dy_buf_offset];
     if(lidy == 0)
@@ -183,6 +183,8 @@ __kernel
     }
     barrier(CLK_LOCAL_MEM_FENCE);
 
+    if(gidx < cols)
+    {
         if(gidy < rows)
         {
             int x =  sdx[lidy][lidx] + 2 * sdx[lidy + 1][lidx] + sdx[lidy + 2][lidx];