int CalcSAD(uchar *im0, uchar *im1, int size)
{
	__m128i s = _mm_setzero_si128();
	__m128i *a = (__m128i *)im0;
	__m128i *b = (__m128i *)im1;
	for (int i = 0; i < size; i+=16) {
		__m128i t = _mm_sad_epu8(*a, *b);
		s = _mm_add_epi32(s, t);
		a += 1;
		b += 1;
	}
	int *p = (int *)&s;
//	printf("SAD = %d\n", p[0]+p[2]);
	return p[0] + p[2];
}