Home | History | Annotate | Download | only in x86

Lines Matching full:temp2

104     __m128i temp1, temp2, temp3, temp4, temp5, temp6;
123 temp2 = _mm_unpacklo_epi8(line3, line4);
127 line1 = _mm_unpacklo_epi16(temp1, temp2);
128 line2 = _mm_unpackhi_epi16(temp1, temp2);
147 temp2 = _mm_unpacklo_epi8(line3, line4);
151 line1 = _mm_unpacklo_epi16(temp1, temp2);
152 line2 = _mm_unpackhi_epi16(temp1, temp2);
157 temp2 = _mm_unpackhi_epi32(line1, line3);
165 p1_16x8 = _mm_unpacklo_epi64(p0_8x16, temp2);
166 p0_16x8 = _mm_unpackhi_epi64(p0_8x16, temp2);
172 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8);
173 temp1 = _mm_add_epi8(temp1, temp2);
175 temp2 = _mm_unpacklo_epi8(temp1, zero);
178 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2);
181 flag1_16x8 = _mm_packs_epi16(temp2, temp1);
185 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8);
186 temp1 = _mm_add_epi8(temp1, temp2);
188 temp2 = _mm_unpacklo_epi8(temp1, zero);
191 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
194 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
200 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8);
201 temp1 = _mm_add_epi8(temp1, temp2);
203 temp2 = _mm_unpacklo_epi8(temp1, zero);
206 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
209 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
216 temp2 = _mm_subs_epu8(q0_16x8, p0_16x8);
217 temp1 = _mm_add_epi8(temp1, temp2);
221 temp2 = _mm_unpacklo_epi8(temp1, zero);
223 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2);
226 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
231 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8);
232 temp1 = _mm_add_epi8(temp1, temp2);
234 temp2 = _mm_unpacklo_epi8(temp1, zero);
236 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
239 flag3_16x8 = _mm_packs_epi16(temp2, temp1);
244 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8);
245 temp1 = _mm_add_epi8(temp1, temp2);
247 temp2 = _mm_unpacklo_epi8(temp1, zero);
249 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
252 flag4_16x8 = _mm_packs_epi16(temp2, temp1);
267 temp2 = _mm_add_epi16(p1_8x16, q0_8x16);
269 temp6 = _mm_add_epi16(temp2, const_val2_16x8);
273 temp2 = _mm_add_epi16(temp6, temp4);
275 q0_16x8_1 = _mm_srai_epi16(temp2, 2);
281 temp2 = _mm_add_epi16(temp5, q2_8x16);
283 q1_16x8_2 = _mm_srai_epi16(temp2, 2);
287 temp2 = _mm_add_epi16(temp4, q2_8x16);
289 temp2 = _mm_add_epi16(temp2, p1_8x16);
293 temp2 = _mm_add_epi16(temp2, temp3);
295 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4));
297 q0_16x8_2 = _mm_srai_epi16(temp2, 3);
301 temp2 = _mm_add_epi16(temp5, const_val2_16x8);
309 temp2 = _mm_add_epi16(temp2, temp4);
311 temp2 = _mm_add_epi16(temp2, temp6);
313 q2_16x8_2 = _mm_srai_epi16(temp2, 3);
327 temp2 = _mm_add_epi16(p1_8x16, q0_8x16);
329 temp6 = _mm_add_epi16(temp2, const_val2_16x8);
333 temp2 = _mm_add_epi16(temp6, temp4);
335 temp2 = _mm_srai_epi16(temp2, 2);
337 q0_16x8_1 = _mm_packus_epi16(q0_16x8_1, temp2);
343 temp2 = _mm_add_epi16(temp5, q2_8x16);
345 temp2 = _mm_srai_epi16(temp2, 2);
347 q1_16x8_2 = _mm_packus_epi16(q1_16x8_2, temp2);
351 temp2 = _mm_add_epi16(temp4, q2_8x16);
353 temp2 = _mm_add_epi16(temp2, p1_8x16);
357 temp2 = _mm_add_epi16(temp2, temp3);
359 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4));
361 temp2 = _mm_srai_epi16(temp2, 3);
363 q0_16x8_2 = _mm_packus_epi16(q0_16x8_2, temp2);
367 temp2 = _mm_add_epi16(temp5, const_val2_16x8);
375 temp2 = _mm_add_epi16(temp2, temp4);
377 temp2 = _mm_add_epi16(temp2, temp6);
379 temp2 = _mm_srai_epi16(temp2, 3);
381 q2_16x8_2 = _mm_packus_epi16(q2_16x8_2, temp2);
424 temp2 = _mm_unpacklo_epi8(p1_16x8, p0_16x8);
428 p3_8x16 = _mm_unpacklo_epi16(temp1, temp2);
429 p2_8x16 = _mm_unpackhi_epi16(temp1, temp2);
452 temp2 = _mm_unpackhi_epi8(p1_16x8, p0_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;
557 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8);
558 temp1 = _mm_add_epi8(temp1, temp2);
560 temp2 = _mm_unpacklo_epi8(temp1, zero);
563 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2);
566 flag1_16x8 = _mm_packs_epi16(temp2, temp1);
570 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8);
571 temp1 = _mm_add_epi8(temp1, temp2);
573 temp2 = _mm_unpacklo_epi8(temp1, zero);
576 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
579 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
585 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8);
586 temp1 = _mm_add_epi8(temp1, temp2);
588 temp2 = _mm_unpacklo_epi8(temp1, zero);
591 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
594 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
601 temp2 = _mm_subs_epu8(q0_16x8, p0_16x8);
602 temp1 = _mm_add_epi8(temp1, temp2);
606 temp2 = _mm_unpacklo_epi8(temp1, zero);
608 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2);
611 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
616 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8);
617 temp1 = _mm_add_epi8(temp1, temp2);
619 temp2 = _mm_unpacklo_epi8(temp1, zero);
621 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
624 flag3_16x8 = _mm_packs_epi16(temp2, temp1);
629 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8);
630 temp1 = _mm_add_epi8(temp1, temp2);
632 temp2 = _mm_unpacklo_epi8(temp1, zero);
634 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
637 flag4_16x8 = _mm_packs_epi16(temp2, temp1);
652 temp2 = _mm_add_epi16(p1_8x16, q0_8x16);
654 temp6 = _mm_add_epi16(temp2, const_val2_16x8);
658 temp2 = _mm_add_epi16(temp6, temp4);
660 q0_16x8_1 = _mm_srai_epi16(temp2, 2);
666 temp2 = _mm_add_epi16(temp5, q2_8x16);
668 q1_16x8_2 = _mm_srai_epi16(temp2, 2);
672 temp2 = _mm_add_epi16(temp4, q2_8x16);
674 temp2 = _mm_add_epi16(temp2, p1_8x16);
678 temp2 = _mm_add_epi16(temp2, temp3);
680 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4));
682 q0_16x8_2 = _mm_srai_epi16(temp2, 3);
686 temp2 = _mm_add_epi16(temp5, const_val2_16x8);
694 temp2 = _mm_add_epi16(temp2, temp4);
696 temp2 = _mm_add_epi16(temp2, temp6);
698 q2_16x8_2 = _mm_srai_epi16(temp2, 3);
712 temp2 = _mm_add_epi16(p1_8x16, q0_8x16);
714 temp6 = _mm_add_epi16(temp2, const_val2_16x8);
718 temp2 = _mm_add_epi16(temp6, temp4);
720 temp2 = _mm_srai_epi16(temp2, 2);
722 q0_16x8_1 = _mm_packus_epi16(q0_16x8_1, temp2);
728 temp2 = _mm_add_epi16(temp5, q2_8x16);
730 temp2 = _mm_srai_epi16(temp2, 2);
732 q1_16x8_2 = _mm_packus_epi16(q1_16x8_2, temp2);
736 temp2 = _mm_add_epi16(temp4, q2_8x16);
738 temp2 = _mm_add_epi16(temp2, p1_8x16);
742 temp2 = _mm_add_epi16(temp2, temp3);
744 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4));
746 temp2 = _mm_srai_epi16(temp2, 3);
748 q0_16x8_2 = _mm_packus_epi16(q0_16x8_2, temp2);
752 temp2 = _mm_add_epi16(temp5, const_val2_16x8);
760 temp2 = _mm_add_epi16(temp2, temp4);
762 temp2 = _mm_add_epi16(temp2, temp6);
764 temp2 = _mm_srai_epi16(temp2, 3);
766 q2_16x8_2 = _mm_packus_epi16(q2_16x8_2, temp2);
1157 __m128i temp1, temp2;
1207 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8);
1208 temp1 = _mm_add_epi8(temp1, temp2);
1210 temp2 = _mm_unpacklo_epi8(temp1, zero);
1213 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2);
1216 flag1_16x8 = _mm_packs_epi16(temp2, temp1);
1221 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8);
1222 temp1 = _mm_add_epi8(temp1, temp2);
1224 temp2 = _mm_unpacklo_epi8(temp1, zero);
1227 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1230 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1236 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8);
1237 temp1 = _mm_add_epi8(temp1, temp2);
1239 temp2 = _mm_unpacklo_epi8(temp1, zero);
1242 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1245 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1252 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8);
1253 temp1 = _mm_add_epi8(temp1, temp2);
1255 temp2 = _mm_unpacklo_epi8(temp1, zero);
1257 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1260 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1263 temp2 = _mm_subs_epi16(zero, temp2);
1266 C_8x16 = _mm_add_epi16(C0_8x16, temp2);
1271 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8);
1272 temp1 = _mm_add_epi8(temp1, temp2);
1274 temp2 = _mm_unpacklo_epi8(temp1, zero);
1276 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1279 flag3_16x8 = _mm_packs_epi16(temp2, temp1);
1282 temp2 = _mm_subs_epi16(zero, temp2);
1285 C_8x16 = _mm_add_epi16(C_8x16, temp2);
1291 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(p1_16x8, zero),
1294 temp1 = _mm_add_epi16(temp1, temp2);
1300 temp2 = _mm_subs_epi16(_mm_unpackhi_epi8(p1_16x8, zero),
1303 temp1 = _mm_add_epi16(temp1, temp2);
1315 temp2 = _mm_add_epi16(_mm_unpackhi_epi8(p0_16x8, zero), in_macro_hi_16x8);
1317 temp1 = _mm_packus_epi16(temp1, temp2);
1320 temp2 = _mm_and_si128(p0_16x8,
1323 temp1 = _mm_add_epi8(temp1, temp2);
1328 temp2 = _mm_sub_epi16(_mm_unpackhi_epi8(q0_16x8, zero), in_macro_hi_16x8);
1330 temp1 = _mm_packus_epi16(temp1, temp2);
1333 temp2 = _mm_and_si128(q0_16x8,
1336 temp1 = _mm_add_epi8(temp1, temp2);
1342 temp2 = _mm_slli_epi16(_mm_unpacklo_epi8(p1_16x8, zero), 1);
1343 //temp2 = _mm_subs_epi16(zero,temp2);
1344 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(p2_16x8, zero), temp2);
1345 temp2 = _mm_add_epi16(temp1, temp2);
1346 in_macro_16x8 = _mm_srai_epi16(temp2, 1);
1350 temp2 = _mm_slli_epi16(_mm_unpackhi_epi8(p1_16x8, zero), 1);
1351 //temp2 = _mm_subs_epi16(zero,temp2);
1352 temp2 = _mm_subs_epi16(_mm_unpackhi_epi8(p2_16x8, zero), temp2);
1353 temp2 = _mm_add_epi16(temp1, temp2);
1354 in_macro_hi_16x8 = _mm_srai_epi16(temp2, 1);
1364 temp2 = _mm_add_epi16(_mm_unpackhi_epi8(p1_16x8, zero), in_macro_hi_16x8);
1366 temp1 = _mm_packus_epi16(temp1, temp2);
1369 temp2 = _mm_and_si128(p1_16x8,
1371 temp1 = _mm_add_epi8(temp1, temp2);
1377 temp2 = _mm_slli_epi16(_mm_unpacklo_epi8(q1_16x8, zero), 1);
1378 //temp2 = _mm_slli_epi16 (temp2, 1);
1379 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(q2_16x8, zero), temp2);
1380 temp2 = _mm_add_epi16(temp1, temp2);
1381 in_macro_16x8 = _mm_srai_epi16(temp2, 1);
1385 temp2 = _mm_slli_epi16(_mm_unpackhi_epi8(q1_16x8, zero), 1);
1386 //temp2 = _mm_slli_epi16 (temp2, 1);
1387 temp2 = _mm_subs_epi16(_mm_unpackhi_epi8(q2_16x8, zero), temp2);
1388 temp2 = _mm_add_epi16(temp1, temp2);
1389 in_macro_hi_16x8 = _mm_srai_epi16(temp2, 1);
1399 temp2 = _mm_add_epi16(_mm_unpackhi_epi8(q1_16x8, zero), in_macro_hi_16x8);
1401 temp1 = _mm_packus_epi16(temp1, temp2);
1404 temp2 = _mm_and_si128(q1_16x8,
1406 temp1 = _mm_add_epi8(temp1, temp2);
1456 __m128i temp1, temp2, temp3, temp4, temp5, temp6;
1475 temp2 = _mm_unpacklo_epi8(line3, line4);
1479 line1 = _mm_unpacklo_epi16(temp1, temp2);
1480 line2 = _mm_unpackhi_epi16(temp1, temp2);
1500 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8);
1501 temp1 = _mm_add_epi8(temp1, temp2);
1503 temp2 = _mm_unpacklo_epi8(temp1, zero);
1506 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2);
1509 flag1_16x8 = _mm_packs_epi16(temp2, temp1);
1513 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8);
1514 temp1 = _mm_add_epi8(temp1, temp2);
1516 temp2 = _mm_unpacklo_epi8(temp1, zero);
1519 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1522 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1528 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8);
1529 temp1 = _mm_add_epi8(temp1, temp2);
1531 temp2 = _mm_unpacklo_epi8(temp1, zero);
1534 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1537 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1544 temp2 = _mm_subs_epu8(q0_16x8, p0_16x8);
1545 temp1 = _mm_add_epi8(temp1, temp2);
1549 temp2 = _mm_unpacklo_epi8(temp1, zero);
1551 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2);
1554 flag2_16x8 = _mm_packs_epi16(temp2, temp1);
1559 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8);
1560 temp1 = _mm_add_epi8(temp1, temp2);
1562 temp2 = _mm_unpacklo_epi8(temp1, zero);
1564 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1567 flag3_16x8 = _mm_packs_epi16(temp2, temp1);
1572 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8);
1573 temp1 = _mm_add_epi8(temp1, temp2);
1575 temp2 = _mm_unpacklo_epi8(temp1, zero);
1577 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1580 flag4_16x8 = _mm_packs_epi16(temp2, temp1);
1595 temp2 = _mm_add_epi16(p1_8x16, q0_8x16);
1597 temp6 = _mm_add_epi16(temp2, const_val2_16x8);
1601 temp2 = _mm_add_epi16(temp6, temp4);
1603 q0_16x8_1 = _mm_srai_epi16(temp2, 2);
1609 temp2 = _mm_add_epi16(temp5, q2_8x16);
1611 q1_16x8_2 = _mm_srai_epi16(temp2, 2);
1615 temp2 = _mm_add_epi16(temp4, q2_8x16);
1617 temp2 = _mm_add_epi16(temp2, p1_8x16);
1621 temp2 = _mm_add_epi16(temp2, temp3);
1623 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4));
1625 q0_16x8_2 = _mm_srai_epi16(temp2, 3);
1629 temp2 = _mm_add_epi16(temp5, const_val2_16x8);
1637 temp2 = _mm_add_epi16(temp2, temp4);
1639 temp2 = _mm_add_epi16(temp2, temp6);
1641 q2_16x8_2 = _mm_srai_epi16(temp2, 3);
1700 temp2 = _mm_unpacklo_epi8(p1_16x8, p0_16x8);
1704 p3_8x16 = _mm_unpacklo_epi16(temp1, temp2);
1705 p2_8x16 = _mm_unpackhi_epi16(temp1, temp2);
1772 __m128i temp1, temp2, temp3, temp4;
1792 temp2 = _mm_unpacklo_epi8(line3, line4);
1796 line1 = _mm_unpacklo_epi16(temp1, temp2);
1797 line2 = _mm_unpackhi_epi16(temp1, temp2);
1802 temp2 = _mm_unpackhi_epi32(line1, line3);
1810 p1_16x8 = _mm_unpacklo_epi64(temp2, zero);
1811 p0_16x8 = _mm_unpackhi_epi64(temp2, zero);
1839 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8);
1840 temp1 = _mm_add_epi8(temp1, temp2);
1842 temp2 = _mm_unpacklo_epi8(temp1, zero);
1843 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2);
1845 flag1_16x8 = _mm_packs_epi16(temp2, zero);
1850 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8);
1851 temp1 = _mm_add_epi8(temp1, temp2);
1853 temp2 = _mm_unpacklo_epi8(temp1, zero);
1854 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1856 flag2_16x8 = _mm_packs_epi16(temp2, zero);
1861 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8);
1862 temp1 = _mm_add_epi8(temp1, temp2);
1864 temp2 = _mm_unpacklo_epi8(temp1, zero);
1865 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1867 flag2_16x8 = _mm_packs_epi16(temp2, zero);
1874 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8);
1875 temp1 = _mm_add_epi8(temp1, temp2);
1877 temp2 = _mm_unpacklo_epi8(temp1, zero);
1878 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1880 flag2_16x8 = _mm_packs_epi16(temp2, zero);
1883 temp2 = _mm_subs_epi16(zero, temp2);
1885 C_8x16 = _mm_add_epi16(C0_8x16, temp2);
1889 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8);
1890 temp1 = _mm_add_epi8(temp1, temp2);
1892 temp2 = _mm_unpacklo_epi8(temp1, zero);
1893 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2);
1895 flag3_16x8 = _mm_packs_epi16(temp2, zero);
1898 temp2 = _mm_subs_epi16(zero, temp2);
1900 C_8x16 = _mm_add_epi16(C_8x16, temp2);
1905 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(p1_16x8, zero),
1908 temp1 = _mm_add_epi16(temp1, temp2);
1941 temp2 = _mm_slli_epi16(_mm_unpacklo_epi8(p1_16x8, zero), 1);
1942 //temp2 = _mm_subs_epi16(zero,temp2);
1943 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(p2_16x8, zero), temp2);
1944 temp2 = _mm_add_epi16(temp1, temp2);
1945 in_macro_16x8 = _mm_srai_epi16(temp2, 1);
1964 temp2 = _mm_slli_epi16(_mm_unpacklo_epi8(q1_16x8, zero), 1);
1965 //temp2 = _mm_slli_epi16 (temp2, 1);
1966 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(q2_16x8, zero), temp2);
1967 temp2 = _mm_add_epi16(temp1, temp2);
1968 in_macro_16x8 = _mm_srai_epi16(temp2, 1);
1985 temp2 = _mm_unpacklo_epi8(p1_16x8, p0_16x8_1);
1989 line7 = _mm_unpacklo_epi16(temp1, temp2);
1990 temp1 = _mm_unpackhi_epi16(temp1, temp2);
1992 temp2 = _mm_unpackhi_epi16(temp3, temp4);
1998 line5 = _mm_unpacklo_epi32(temp1, temp2);
2000 line7 = _mm_unpackhi_epi32(temp1, temp2);