| 197 | | __m128i r1 = _mm_unpacklo_epi8(chroma_l, tmp_y), |
|---|
| 198 | | r3 = _mm_unpacklo_epi8(chroma_h, tmp_y3), |
|---|
| 199 | | r2 = _mm_unpackhi_epi8(chroma_l, tmp_y), |
|---|
| 200 | | r4 = _mm_unpackhi_epi8(chroma_h, tmp_y3), |
|---|
| 201 | | r5 = _mm_unpacklo_epi8(chroma_l, tmp_y2), |
|---|
| 202 | | r7 = _mm_unpacklo_epi8(chroma_h, tmp_y4), |
|---|
| 203 | | r6 = _mm_unpackhi_epi8(chroma_l, tmp_y2), |
|---|
| 204 | | r8 = _mm_unpackhi_epi8(chroma_h, tmp_y4); |
|---|
| | 197 | _mm_stream_si128(&ov[x4], _mm_unpacklo_epi8(chroma_l, tmp_y)); |
|---|
| | 198 | _mm_stream_si128(&ov[x4+1], _mm_unpackhi_epi8(chroma_l, tmp_y)); |
|---|
| | 199 | _mm_stream_si128(&ov[x4+2], _mm_unpacklo_epi8(chroma_h, tmp_y3)); |
|---|
| | 200 | _mm_stream_si128(&ov[x4+3], _mm_unpackhi_epi8(chroma_h, tmp_y3)); |
|---|
| 206 | | _mm_stream_si128(&ov[x4], r1); |
|---|
| 207 | | _mm_stream_si128(&ov[x4+1], r2); |
|---|
| 208 | | _mm_stream_si128(&ov[x4+2], r3); |
|---|
| 209 | | _mm_stream_si128(&ov[x4+3], r4); |
|---|
| 210 | | |
|---|
| 211 | | _mm_stream_si128(&ov2[x4], r5); |
|---|
| 212 | | _mm_stream_si128(&ov2[x4+1],r6); |
|---|
| 213 | | _mm_stream_si128(&ov2[x4+2],r7); |
|---|
| 214 | | _mm_stream_si128(&ov2[x4+3],r8); |
|---|
| | 202 | _mm_stream_si128(&ov2[x4], _mm_unpacklo_epi8(chroma_l, tmp_y2)); |
|---|
| | 203 | _mm_stream_si128(&ov2[x4+1],_mm_unpackhi_epi8(chroma_l, tmp_y2)); |
|---|
| | 204 | _mm_stream_si128(&ov2[x4+2],_mm_unpacklo_epi8(chroma_h, tmp_y4)); |
|---|
| | 205 | _mm_stream_si128(&ov2[x4+3],_mm_unpackhi_epi8(chroma_h, tmp_y4)); |
|---|