fix some mismatch on cpu device running OCL

modules/ocl/src/opencl/stereobm.cl

@@ -228,7 +228,7 @@ __kernel void stereoKernel(__global unsigned char *left, __global unsigned char
     int X = get_group_id(0) * BLOCK_W + get_local_id(0) + maxdisp + radius;
     // int Y = get_group_id(1) * ROWSperTHREAD + radius;
 
-    #define Y (get_group_id(1) * ROWSperTHREAD + radius)
+#define Y (get_group_id(1) * ROWSperTHREAD + radius)
 
     volatile __global unsigned int* minSSDImage = cminSSDImage + X + Y * cminSSD_step;
     __global unsigned char* disparImage = disp + X + Y * disp_step;
@@ -251,9 +251,9 @@ __kernel void stereoKernel(__global unsigned char *left, __global unsigned char
 
         barrier(CLK_LOCAL_MEM_FENCE); //before MinSSD function
 
+        uint2 minSSD = MinSSD(col_ssd_cache + get_local_id(0), col_ssd, radius);
         if (X < cwidth - radius && Y < cheight - radius)
         {
-            uint2 minSSD = MinSSD(col_ssd_cache + get_local_id(0), col_ssd, radius);
             if (minSSD.x < minSSDImage[0])
             {
                 disparImage[0] = (unsigned char)(d + minSSD.y);
@@ -264,7 +264,7 @@ __kernel void stereoKernel(__global unsigned char *left, __global unsigned char
         for(int row = 1; row < end_row; row++)
         {
             int idx1 = y_tex * img_step + x_tex;
-            int idx2 = (y_tex + (2 * radius + 1)) * img_step + x_tex;
+            int idx2 = min(y_tex + (2 * radius + 1), cheight - 1) * img_step + x_tex;
 
             barrier(CLK_GLOBAL_MEM_FENCE);
             barrier(CLK_LOCAL_MEM_FENCE);
@@ -278,10 +278,10 @@ __kernel void stereoKernel(__global unsigned char *left, __global unsigned char
 
             barrier(CLK_LOCAL_MEM_FENCE);
 
+            uint2 minSSD = MinSSD(col_ssd_cache + get_local_id(0), col_ssd, radius);
             if (X < cwidth - radius && row < cheight - radius - Y)
             {
                 int idx = row * cminSSD_step;
-                uint2 minSSD = MinSSD(col_ssd_cache + get_local_id(0), col_ssd, radius);
                 if (minSSD.x < minSSDImage[idx])
                 {
                     disparImage[disp_step * row] = (unsigned char)(d + minSSD.y);
@@ -378,8 +378,8 @@ __kernel void textureness_kernel(__global unsigned char *disp, int disp_rows, in
     int beg_row = group_id_y * RpT;
     int end_row = min(beg_row + RpT, disp_rows);
 
- //   if (x < disp_cols)
- //   {
+//   if (x < disp_cols)
+//   {
     int y = beg_row;
 
     float sum = 0;