void X264_Decode_Convert_gbr24p_to_packed32(const uint8_t** src, uint8_t *dst, int width, int height) {
    int arrlen = width * height * 4;

    uint32_t *dp = (uint32_t*)dst;
    const uint8_t* __restrict gp = src[0];
    const uint8_t* __restrict bp = src[1];
    const uint8_t* __restrict rp = src[2];
    for (int i = 0; i < arrlen; i++) {
    {
        *(dp++) = (0xff<<24) | *rp++ << 16 | *gp++ << 8 | *bp++;
    }
}