Lines Matching full:temp1
104 __m128i temp1, temp2, temp3, temp4, temp5, temp6;
122 temp1 = _mm_unpacklo_epi8(line1, line2);
127 line1 = _mm_unpacklo_epi16(temp1, temp2);
128 line2 = _mm_unpackhi_epi16(temp1, temp2);
146 temp1 = _mm_unpacklo_epi8(line1, line2);
151 line1 = _mm_unpacklo_epi16(temp1, temp2);
152 line2 = _mm_unpackhi_epi16(temp1, temp2);
156 temp1 = _mm_unpacklo_epi32(line1, line3);
161 p3_16x8 = _mm_unpacklo_epi64(p1_8x16, temp1);
162 p2_16x8 = _mm_unpackhi_epi64(p1_8x16, temp1);
171 temp1 = _mm_subs_epu8(q0_16x8, p0_16x8);
173 temp1 = _mm_add_epi8(temp1, temp2);
175 temp2 = _mm_unpacklo_epi8(temp1, zero);
176 temp1 = _mm_unpackhi_epi8(temp1, zero);
179 temp1 = _mm_cmpgt_epi16(Alpha_8x16, temp1);
181 flag1_16x8 = _mm_packs_epi16(temp2, temp1);
184 temp1 = _mm_subs_epu8(q0_16x8, q1_16x8);
186 temp1 = _mm_add_epi8(temp1, temp2);
188 temp2 = _mm_unpacklo_epi8(temp1, zero);
189 temp1 = _mm_unpackhi_epi8(temp1, zero);
192 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
194 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
199 temp1 = _mm_subs_epu8(p0_16x8, p1_16x8);
201 temp1 = _mm_add_epi8(temp1, temp2);
203 temp2 = _mm_unpacklo_epi8(temp1, zero);
204 temp1 = _mm_unpackhi_epi8(temp1, zero);
207 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
209 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
215 temp1 = _mm_subs_epu8(p0_16x8, q0_16x8);
217 temp1 = _mm_add_epi8(temp1, temp2);
221 temp2 = _mm_unpacklo_epi8(temp1, zero);
222 temp1 = _mm_unpackhi_epi8(temp1, zero);
224 temp1 = _mm_cmpgt_epi16(Alpha_8x16, temp1);
226 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
230 temp1 = _mm_subs_epu8(p0_16x8, p2_16x8);
232 temp1 = _mm_add_epi8(temp1, temp2);
234 temp2 = _mm_unpacklo_epi8(temp1, zero);
235 temp1 = _mm_unpackhi_epi8(temp1, zero);
237 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
239 flag3_16x8 = _mm_packs_epi16(temp2, temp1);
243 temp1 = _mm_subs_epu8(q0_16x8, q2_16x8);
245 temp1 = _mm_add_epi8(temp1, temp2);
247 temp2 = _mm_unpacklo_epi8(temp1, zero);
248 temp1 = _mm_unpackhi_epi8(temp1, zero);
250 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
252 flag4_16x8 = _mm_packs_epi16(temp2, temp1);
266 temp1 = _mm_add_epi16(p0_8x16, q1_8x16);
268 temp5 = _mm_add_epi16(temp1, const_val2_16x8);
272 temp1 = _mm_add_epi16(temp5, temp3);
274 p0_16x8_1 = _mm_srai_epi16(temp1, 2);
280 temp1 = _mm_add_epi16(temp6, p2_8x16);
282 p1_16x8_2 = _mm_srai_epi16(temp1, 2);
286 temp1 = _mm_add_epi16(temp3, p2_8x16);
288 temp1 = _mm_add_epi16(temp1, q1_8x16);
292 temp1 = _mm_add_epi16(temp1, temp3);
294 temp1 = _mm_add_epi16(temp1, _mm_set1_epi16(4));
296 p0_16x8_2 = _mm_srai_epi16(temp1, 3);
300 temp1 = _mm_add_epi16(temp6, const_val2_16x8);
308 temp1 = _mm_add_epi16(temp1, temp3);
310 temp1 = _mm_add_epi16(temp1, temp5);
312 p2_16x8_2 = _mm_srai_epi16(temp1, 3);
326 temp1 = _mm_add_epi16(p0_8x16, q1_8x16);
328 temp5 = _mm_add_epi16(temp1, const_val2_16x8);
332 temp1 = _mm_add_epi16(temp5, temp3);
334 temp1 = _mm_srai_epi16(temp1, 2);
336 p0_16x8_1 = _mm_packus_epi16(p0_16x8_1, temp1);
342 temp1 = _mm_add_epi16(temp6, p2_8x16);
344 temp1 = _mm_srai_epi16(temp1, 2);
346 p1_16x8_2 = _mm_packus_epi16(p1_16x8_2, temp1);
350 temp1 = _mm_add_epi16(temp3, p2_8x16);
352 temp1 = _mm_add_epi16(temp1, q1_8x16);
356 temp1 = _mm_add_epi16(temp1, temp3);
358 temp1 = _mm_add_epi16(temp1, _mm_set1_epi16(4));
360 temp1 = _mm_srai_epi16(temp1, 3);
362 temp1);
366 temp1 = _mm_add_epi16(temp6, const_val2_16x8);
374 temp1 = _mm_add_epi16(temp1, temp3);
376 temp1 = _mm_add_epi16(temp1, temp5);
378 temp1 = _mm_srai_epi16(temp1, 3);
380 p2_16x8_2 = _mm_packus_epi16(p2_16x8_2, temp1);
423 temp1 = _mm_unpacklo_epi8(p3_16x8, p2_16x8);
428 p3_8x16 = _mm_unpacklo_epi16(temp1, temp2);
429 p2_8x16 = _mm_unpackhi_epi16(temp1, temp2);
451 temp1 = _mm_unpackhi_epi8(p3_16x8, p2_16x8);
456 p3_8x16 = _mm_unpacklo_epi16(temp1, temp2);
457 p2_8x16 = _mm_unpackhi_epi16(temp1, temp2);
528 __m128i temp1, temp2, temp3, temp4, temp5, temp6;
556 temp1 = _mm_subs_epu8(q0_16x8, p0_16x8);
558 temp1 = _mm_add_epi8(temp1, temp2);
560 temp2 = _mm_unpacklo_epi8(temp1, zero);
561 temp1 = _mm_unpackhi_epi8(temp1, zero);
564 temp1 = _mm_cmpgt_epi16(Alpha_8x16, temp1);
566 flag1_16x8 = _mm_packs_epi16(temp2, temp1);
569 temp1 = _mm_subs_epu8(q0_16x8, q1_16x8);
571 temp1 = _mm_add_epi8(temp1, temp2);
573 temp2 = _mm_unpacklo_epi8(temp1, zero);
574 temp1 = _mm_unpackhi_epi8(temp1, zero);
577 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
579 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
584 temp1 = _mm_subs_epu8(p0_16x8, p1_16x8);
586 temp1 = _mm_add_epi8(temp1, temp2);
588 temp2 = _mm_unpacklo_epi8(temp1, zero);
589 temp1 = _mm_unpackhi_epi8(temp1, zero);
592 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
594 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
600 temp1 = _mm_subs_epu8(p0_16x8, q0_16x8);
602 temp1 = _mm_add_epi8(temp1, temp2);
606 temp2 = _mm_unpacklo_epi8(temp1, zero);
607 temp1 = _mm_unpackhi_epi8(temp1, zero);
609 temp1 = _mm_cmpgt_epi16(Alpha_8x16, temp1);
611 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
615 temp1 = _mm_subs_epu8(p0_16x8, p2_16x8);
617 temp1 = _mm_add_epi8(temp1, temp2);
619 temp2 = _mm_unpacklo_epi8(temp1, zero);
620 temp1 = _mm_unpackhi_epi8(temp1, zero);
622 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
624 flag3_16x8 = _mm_packs_epi16(temp2, temp1);
628 temp1 = _mm_subs_epu8(q0_16x8, q2_16x8);
630 temp1 = _mm_add_epi8(temp1, temp2);
632 temp2 = _mm_unpacklo_epi8(temp1, zero);
633 temp1 = _mm_unpackhi_epi8(temp1, zero);
635 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
637 flag4_16x8 = _mm_packs_epi16(temp2, temp1);
651 temp1 = _mm_add_epi16(p0_8x16, q1_8x16);
653 temp5 = _mm_add_epi16(temp1, const_val2_16x8);
657 temp1 = _mm_add_epi16(temp5, temp3);
659 p0_16x8_1 = _mm_srai_epi16(temp1, 2);
665 temp1 = _mm_add_epi16(temp6, p2_8x16);
667 p1_16x8_2 = _mm_srai_epi16(temp1, 2);
671 temp1 = _mm_add_epi16(temp3, p2_8x16);
673 temp1 = _mm_add_epi16(temp1, q1_8x16);
677 temp1 = _mm_add_epi16(temp1, temp3);
679 temp1 = _mm_add_epi16(temp1, _mm_set1_epi16(4));
681 p0_16x8_2 = _mm_srai_epi16(temp1, 3);
685 temp1 = _mm_add_epi16(temp6, const_val2_16x8);
693 temp1 = _mm_add_epi16(temp1, temp3);
695 temp1 = _mm_add_epi16(temp1, temp5);
697 p2_16x8_2 = _mm_srai_epi16(temp1, 3);
711 temp1 = _mm_add_epi16(p0_8x16, q1_8x16);
713 temp5 = _mm_add_epi16(temp1, const_val2_16x8);
717 temp1 = _mm_add_epi16(temp5, temp3);
719 temp1 = _mm_srai_epi16(temp1, 2);
721 p0_16x8_1 = _mm_packus_epi16(p0_16x8_1, temp1);
727 temp1 = _mm_add_epi16(temp6, p2_8x16);
729 temp1 = _mm_srai_epi16(temp1, 2);
731 p1_16x8_2 = _mm_packus_epi16(p1_16x8_2, temp1);
735 temp1 = _mm_add_epi16(temp3, p2_8x16);
737 temp1 = _mm_add_epi16(temp1, q1_8x16);
741 temp1 = _mm_add_epi16(temp1, temp3);
743 temp1 = _mm_add_epi16(temp1, _mm_set1_epi16(4));
745 temp1temp1, 3);
747 p0_16x8_2 = _mm_packus_epi16(p0_16x8_2, temp1);
751 temp1 = _mm_add_epi16(temp6, const_val2_16x8);
759 temp1 = _mm_add_epi16(temp1, temp3);
761 temp1 = _mm_add_epi16(temp1, temp5);
763 temp1 = _mm_srai_epi16(temp1, 3);
765 p2_16x8_2 = _mm_packus_epi16(p2_16x8_2, temp1);
865 temp1;
1000 temp1 = _mm_and_si128(lined,
1002 temp1 = _mm_add_epi16(temp1, in_macrotemp);
1003 //temp1= _mm_packus_epi16 (temp1, zero);
1041 lined = temp1;
1157 __m128i temp1, temp2;
1206 temp1 = _mm_subs_epu8(q0_16x8, p0_16x8);
1208 temp1 = _mm_add_epi8(temp1, temp2);
1210 temp2 = _mm_unpacklo_epi8(temp1, zero);
1211 temp1 = _mm_unpackhi_epi8(temp1, zero);
1214 temp1 = _mm_cmpgt_epi16(Alpha_8x16, temp1);
1216 flag1_16x8 = _mm_packs_epi16(temp2, temp1);
1220 temp1 = _mm_subs_epu8(q0_16x8, q1_16x8);
1222 temp1 = _mm_add_epi8(temp1, temp2);
1224 temp2 = _mm_unpacklo_epi8(temp1, zero);
1225 temp1 = _mm_unpackhi_epi8(temp1, zero);
1228 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
1230 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1235 temp1 = _mm_subs_epu8(p0_16x8, p1_16x8);
1237 temp1 = _mm_add_epi8(temp1, temp2);
1239 temp2 = _mm_unpacklo_epi8(temp1, zero);
1240 temp1 = _mm_unpackhi_epi8(temp1, zero);
1243 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
1245 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1251 temp1 = _mm_subs_epu8(p0_16x8, p2_16x8);
1253 temp1 = _mm_add_epi8(temp1, temp2);
1255 temp2 = _mm_unpacklo_epi8(temp1, zero);
1256 temp1 = _mm_unpackhi_epi8(temp1, zero);
1258 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
1260 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1264 temp1 = _mm_subs_epi16(zero, temp1);
1267 C_hi_8x16 = _mm_add_epi16(C0_hi_8x16, temp1);
1270 temp1 = _mm_subs_epu8(q0_16x8, q2_16x8);
1272 temp1 = _mm_add_epi8(temp1, temp2);
1274 temp2 = _mm_unpacklo_epi8(temp1, zero);
1275 temp1 = _mm_unpackhi_epi8(temp1, zero);
1277 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
1279 flag3_16x8 = _mm_packs_epi16(temp2, temp1);
1283 temp1 = _mm_subs_epi16(zero, temp1);
1286 C_hi_8x16 = _mm_add_epi16(C_hi_8x16, temp1);
1289 temp1 = _mm_subs_epi16(_mm_unpacklo_epi8(q0_16x8, zero),
1293 temp1 = _mm_slli_epi16(temp1, 2);
1294 temp1 = _mm_add_epi16(temp1, temp2);
1295 temp1 = _mm_add_epi16(temp1, const_val4_8x16);
1296 in_macro_16x8 = _mm_srai_epi16(temp1, 3);
1298 temp1 = _mm_subs_epi16(_mm_unpackhi_epi8(q0_16x8, zero),
1302 temp1 = _mm_slli_epi16(temp1, 2);
1303 temp1 = _mm_add_epi16(temp1, temp2);
1304 temp1 = _mm_add_epi16(temp1, const_val4_8x16);
1305 in_macro_hi_16x8 = _mm_srai_epi16(temp1, 3);
1314 temp1 = _mm_add_epi16(_mm_unpacklo_epi8(p0_16x8, zero), in_macro_16x8);
1317 temp1 = _mm_packus_epi16(temp1, temp2);
1319 temp1 = _mm_and_si128(temp1, flag1_16x8);
1323 temp1 = _mm_add_epi8(temp1, temp2);
1325 _mm_storeu_si128((__m128i *)(pu1_HorzPixel + i16_posP0), temp1);
1327 temp1 = _mm_sub_epi16(_mm_unpacklo_epi8(q0_16x8, zero), in_macro_16x8);
1330 temp1 = _mm_packus_epi16(temp1, temp2);
1332 temp1 = _mm_and_si128(temp1, flag1_16x8);
1336 temp1 = _mm_add_epi8(temp1, temp2);
1337 _mm_storeu_si128((__m128i *)(pu1_src), temp1);
1340 temp1 = _mm_avg_epu16(_mm_unpacklo_epi8(q0_16x8, zero),
1345 temp2 = _mm_add_epi16(temp1, temp2);
1348 temp1 = _mm_avg_epu16(_mm_unpackhi_epi8(q0_16x8, zero),
1353 temp2 = _mm_add_epi16(temp1, temp2);
1363 temp1 = _mm_add_epi16(_mm_unpacklo_epi8(p1_16x8, zero), in_macro_16x8);
1366 temp1 = _mm_packus_epi16(temp1, temp2);
1368 temp1 = _mm_and_si128(temp1, flag2_16x8);
1371 temp1 = _mm_add_epi8(temp1, temp2);
1372 _mm_storeu_si128((__m128i *)(pu1_HorzPixel + i16_posP1), temp1);
1375 temp1 = _mm_avg_epu16(_mm_unpacklo_epi8(q0_16x8, zero),
1380 temp2 = _mm_add_epi16(temp1, temp2);
1383 temp1 = _mm_avg_epu16(_mm_unpackhi_epi8(q0_16x8, zero),
1388 temp2 = _mm_add_epi16(temp1, temp2);
1398 temp1 = _mm_add_epi16(_mm_unpacklo_epi8(q1_16x8, zero), in_macro_16x8);
1401 temp1 = _mm_packus_epi16(temp1, temp2);
1403 temp1 = _mm_and_si128(temp1, flag3_16x8);
1406 temp1 = _mm_add_epi8(temp1, temp2);
1408 _mm_storeu_si128((__m128i *)(pu1_src + i16_posQ1), temp1);
1456 __m128i temp1, temp2, temp3, temp4, temp5, temp6;
1474 temp1 = _mm_unpacklo_epi8(line1, line2);
1479 line1 = _mm_unpacklo_epi16(temp1, temp2);
1480 line2 = _mm_unpackhi_epi16(temp1, temp2);
1499 temp1 = _mm_subs_epu8(q0_16x8, p0_16x8);
1501 temp1 = _mm_add_epi8(temp1, temp2);
1503 temp2 = _mm_unpacklo_epi8(temp1, zero);
1504 temp1 = _mm_unpackhi_epi8(temp1, zero);
1507 temp1 = _mm_cmpgt_epi16(Alpha_8x16, temp1);
1509 flag1_16x8 = _mm_packs_epi16(temp2, temp1);
1512 temp1 = _mm_subs_epu8(q0_16x8, q1_16x8);
1514 temp1 = _mm_add_epi8(temp1, temp2);
1516 temp2 = _mm_unpacklo_epi8(temp1, zero);
1517 temp1 = _mm_unpackhi_epi8(temp1, zero);
1520 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
1522 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1527 temp1 = _mm_subs_epu8(p0_16x8, p1_16x8);
1529 temp1 = _mm_add_epi8(temp1, temp2);
1531 temp2 = _mm_unpacklo_epi8(temp1, zero);
1532 temp1 = _mm_unpackhi_epi8(temp1, zero);
1535 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
1537 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1543 temp1 = _mm_subs_epu8(p0_16x8, q0_16x8);
1545 temp1 = _mm_add_epi8(temp1, temp2);
1549 temp2 = _mm_unpacklo_epi8(temp1, zero);
1550 temp1 = _mm_unpackhi_epi8(temp1, zero);
1552 temp1 = _mm_cmpgt_epi16(Alpha_8x16, temp1);
1554 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1558 temp1 = _mm_subs_epu8(p0_16x8, p2_16x8);
1560 temp1 = _mm_add_epi8(temp1, temp2);
1562 temp2 = _mm_unpacklo_epi8(temp1, zero);
1563 temp1 = _mm_unpackhi_epi8(temp1, zero);
1565 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
1567 flag3_16x8 = _mm_packs_epi16(temp2, temp1);
1571 temp1 = _mm_subs_epu8(q0_16x8, q2_16x8);
1573 temp1 = _mm_add_epi8(temp1, temp2);
1575 temp2 = _mm_unpacklo_epi8(temp1, zero);
1576 temp1 = _mm_unpackhi_epi8(temp1, zero);
1578 temp1 = _mm_cmpgt_epi16(Beta_8x16, temp1);
1580 flag4_16x8 = _mm_packs_epi16(temp2, temp1);
1594 temp1 = _mm_add_epi16(p0_8x16, q1_8x16);
1596 temp5 = _mm_add_epi16(temp1, const_val2_16x8);
1600 temp1 = _mm_add_epi16(temp5, temp3);
1602 p0_16x8_1 = _mm_srai_epi16(temp1, 2);
1608 temp1 = _mm_add_epi16(temp6, p2_8x16);
1610 p1_16x8_2 = _mm_srai_epi16(temp1, 2);
1614 temp1 = _mm_add_epi16(temp3, p2_8x16);
1616 temp1 = _mm_add_epi16(temp1, q1_8x16);
1620 temp1 = _mm_add_epi16(temp1, temp3);
1622 temp1 = _mm_add_epi16(temp1, _mm_set1_epi16(4));
1624 p0_16x8_2 = _mm_srai_epi16(temp1, 3);
1628 temp1 = _mm_add_epi16(temp6, const_val2_16x8);
1636 temp1 = _mm_add_epi16(temp1, temp3);
1638 temp1 = _mm_add_epi16(temp1, temp5);
1640 p2_16x8_2 = _mm_srai_epi16(temp1, 3);
1699 temp1 = _mm_unpacklo_epi8(p3_16x8, p2_16x8);
1704 p3_8x16 = _mm_unpacklo_epi16(temp1, temp2);
1705 p2_8x16 = _mm_unpackhi_epi16(temp1, temp2);
1772 __m128i temp1, temp2, temp3, temp4;
1791 temp1 = _mm_unpacklo_epi8(line1, line2);
1796 line1 = _mm_unpacklo_epi16(temp1, temp2);
1797 line2 = _mm_unpackhi_epi16(temp1, temp2);
1801 temp1 = _mm_unpacklo_epi32(line1, line3);
1806 p3_16x8 = _mm_unpacklo_epi64(temp1, zero);
1807 p2_16x8 = _mm_unpackhi_epi64(temp1, zero);
1838 temp1 = _mm_subs_epu8(q0_16x8, p0_16x8);
1840 temp1 = _mm_add_epi8(temp1, temp2);
1842 temp2 = _mm_unpacklo_epi8(temp1, zero);
1849 temp1 = _mm_subs_epu8(q0_16x8, q1_16x8);
1851 temp1 = _mm_add_epi8(temp1, temp2);
1853 temp2 = _mm_unpacklo_epi8(temp1, zero);
1860 temp1 = _mm_subs_epu8(p0_16x8, p1_16x8);
1862 temp1 = _mm_add_epi8(temp1, temp2);
1864 temp2 = _mm_unpacklo_epi8(temp1, zero);
1873 temp1 = _mm_subs_epu8(p0_16x8, p2_16x8);
1875 temp1 = _mm_add_epi8(temp1, temp2);
1877 temp2 = _mm_unpacklo_epi8(temp1, zero);
1888 temp1 = _mm_subs_epu8(q0_16x8, q2_16x8);
1890 temp1 = _mm_add_epi8(temp1, temp2);
1892 temp2 = _mm_unpacklo_epi8(temp1, zero);
1903 temp1 = _mm_subs_epi16(_mm_unpacklo_epi8(q0_16x8, zero),
1907 temp1 = _mm_slli_epi16(temp1, 2);
1908 temp1 = _mm_add_epi16(temp1, temp2);
1909 temp1 = _mm_add_epi16(temp1, const_val4_8x16);
1910 in_macro_16x8 = _mm_srai_epi16(temp1, 3);
1917 temp1 = _mm_add_epi16(_mm_unpacklo_epi8(p0_16x8, zero), in_macro_16x8);
1919 temp1 = _mm_packus_epi16(temp1, zero);
1921 p0_16x8_1 = _mm_and_si128(temp1, flag1_16x8);
1928 temp1 = _mm_sub_epi16(_mm_unpacklo_epi8(q0_16x8, zero), in_macro_16x8);
1930 temp1 = _mm_packus_epi16(temp1, zero);
1932 q0_16x8_1 = _mm_and_si128(temp1, flag1_16x8);
1939 temp1 = _mm_avg_epu16(_mm_unpacklo_epi8(q0_16x8, zero),
1944 temp2 = _mm_add_epi16(temp1, temp2);
1952 temp1 = _mm_add_epi16(_mm_unpacklo_epi8(p1_16x8, zero), in_macro_16x8);
1954 temp1 = _mm_packus_epi16(temp1, zero);
1956 p1_16x8_1 = _mm_and_si128(temp1, flag2_16x8);
1962 temp1 = _mm_avg_epu16(_mm_unpacklo_epi8(q0_16x8, zero),
1967 temp2 = _mm_add_epi16(temp1, temp2);
1974 temp1 = _mm_add_epi16(_mm_unpacklo_epi8(q1_16x8, zero), in_macro_16x8);
1977 temp1 = _mm_packus_epi16(temp1, zero);
1979 q1_16x8_1 = _mm_and_si128(temp1, flag3_16x8);
1984 temp1 = _mm_unpacklo_epi8(p3_16x8, p2_16x8);
1989 line7 = _mm_unpacklo_epi16(temp1, temp2);
1990 temp1 = _mm_unpackhi_epi16(temp1, temp2);
1998 line5 = _mm_unpacklo_epi32(temp1, temp2);
2000 line7 = _mm_unpackhi_epi32(temp1, temp2);
