int bw = blockDim.x;  
int bh = blockDim.y;

int tx = threadIdx.x%bw;  
int ty = threadIdx.y%bh;

\_\_shared\_\_ uchar2 ys0\[\]\[\];  
\_\_shared\_\_ uchar2 ys1\[\]\[\];  
\_\_shared\_\_ uchar2 uvs\[\]\[\];

ys0\[ty\]\[tx\] = y0y1;  
ys1\[ty\]\[tx\] = y2y3;  
uvs\[ty\]\[tx\] = uv;

\_\_syncthreads();  
if (threadIdx.x ==  && threadIdx.y == ) {  
    for (int j = ; j != bh; ++j) {  
        uchar2\* py0 = (uchar2\*)(pDst + (iy + j) \*  \* nPitch + ix \* );  
        uchar2\* py1 = (uchar2\*)(pDst + ((iy + j) \* +) \* nPitch + ix \* );  
        uchar2\* puv = (uchar2\*)(pDstUv + (iy + j)\*nWidth + ix \* );  
        for (int i = ; i != bw; ++i) {  
            \*py0++ = ys0\[j\]\[i\];  
            \*py1++ = ys1\[j\]\[i\];  
            //\*puv++ = uvs\[j\]\[i\];  
        }

    }  
}