ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3,
res0, res1, res2, res3);
ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3);
- res0 = CLIP_SW_0_255(res0);
- res1 = CLIP_SW_0_255(res1);
- res2 = CLIP_SW_0_255(res2);
- res3 = CLIP_SW_0_255(res3);
+ CLIP_SW4_0_255(res0, res1, res2, res3);
VSHF_B2_SB(res0, res1, res2, res3, mask, mask, dest0, dest1);
- ST4x4_UB(dest0, dest1, 0, 1, 0, 1, dst, stride);
+ ST_W2(dest0, 0, 1, dst, stride);
+ ST_W2(dest1, 0, 1, dst + 2 * stride, stride);
memset(input, 0, 4 * 4 * sizeof(*input));
}
ADD4(res0, vec, res1, vec, res2, vec, res3, vec, res0, res1, res2, res3);
CLIP_SH4_0_255(res0, res1, res2, res3);
VSHF_B2_SB(res0, res1, res2, res3, mask, mask, dest0, dest1);
- ST4x4_UB(dest0, dest1, 0, 1, 0, 1, dst, stride);
+ ST_W2(dest0, 0, 1, dst, stride);
+ ST_W2(dest1, 0, 1, dst + 2 * stride, stride);
in_dc[0] = 0;
}