Home | History | Annotate | Download | only in source

Lines Matching full:xmm1

130     "movdqa    %%xmm0,%%xmm1                   \n"
132 "punpckhwd %%xmm1,%%xmm1 \n"
134 "por %%xmm5,%%xmm1 \n"
136 "movdqa %%xmm1,0x10(%1) \n"
146 , "xmm0", "xmm1", "xmm5"
254 "movdqu 0x10(%0),%%xmm1 \n"
258 "palignr $0x8,%%xmm1,%%xmm2 \n"
261 "palignr $0xc,%%xmm0,%%xmm1 \n"
265 "pshufb %%xmm4,%%xmm1 \n"
267 "por %%xmm5,%%xmm1 \n"
270 "movdqa %%xmm1,0x10(%1) \n"
282 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
295 "movdqu 0x10(%0),%%xmm1 \n"
299 "palignr $0x8,%%xmm1,%%xmm2 \n"
302 "palignr $0xc,%%xmm0,%%xmm1 \n"
306 "pshufb %%xmm4,%%xmm1 \n"
308 "por %%xmm5,%%xmm1 \n"
311 "movdqa %%xmm1,0x10(%1) \n"
323 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
348 "movdqa %%xmm0,%%xmm1 \n"
350 "pand %%xmm3,%%xmm1 \n"
352 "pmulhuw %%xmm5,%%xmm1 \n"
354 "psllw $0x8,%%xmm1 \n"
355 "por %%xmm2,%%xmm1 \n"
359 "movdqa %%xmm1,%%xmm2 \n"
360 "punpcklbw %%xmm0,%%xmm1 \n"
362 "movdqa %%xmm1,(%1,%0,2) \n"
373 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
397 "movdqa %%xmm0,%%xmm1 \n"
399 "psllw $0x1,%%xmm1 \n"
401 "pand %%xmm3,%%xmm1 \n"
403 "pmulhuw %%xmm5,%%xmm1 \n"
404 "psllw $0x8,%%xmm1 \n"
405 "por %%xmm2,%%xmm1 \n"
412 "movdqa %%xmm1,%%xmm2 \n"
413 "punpcklbw %%xmm0,%%xmm1 \n"
415 "movdqa %%xmm1,(%1,%0,2) \n"
426 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
446 "movdqa %%xmm0,%%xmm1 \n"
448 "psllw $0x4,%%xmm1 \n"
450 "por %%xmm1,%%xmm0 \n"
452 "movdqa %%xmm0,%%xmm1 \n"
454 "punpckhbw %%xmm2,%%xmm1 \n"
456 "movdqa %%xmm1,0x10(%1,%0,2) \n"
466 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
477 "movdqa 0x10(%0),%%xmm1 \n"
482 "pshufb %%xmm6,%%xmm1 \n"
485 "movdqa %%xmm1,%%xmm4 \n"
486 "psrldq $0x4,%%xmm1 \n"
492 "por %%xmm5,%%xmm1 \n"
496 "movdqa %%xmm1,0x10(%1) \n"
507 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
518 "movdqa 0x10(%0),%%xmm1 \n"
523 "pshufb %%xmm6,%%xmm1 \n"
526 "movdqa %%xmm1,%%xmm4 \n"
527 "psrldq $0x4,%%xmm1 \n"
533 "por %%xmm5,%%xmm1 \n"
537 "movdqa %%xmm1,0x10(%1) \n"
548 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
565 "movdqa %%xmm0,%%xmm1 \n"
568 "psrld $0x3,%%xmm1 \n"
571 "pand %%xmm3,%%xmm1 \n"
574 "por %%xmm2,%%xmm1 \n"
575 "por %%xmm1,%%xmm0 \n"
588 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
606 "movdqa %%xmm0,%%xmm1 \n"
610 "psrld $0x3,%%xmm1 \n"
614 "pand %%xmm4,%%xmm1 \n"
617 "por %%xmm1,%%xmm0 \n"
632 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
646 "movdqa %%xmm0,%%xmm1 \n"
648 "pand %%xmm4,%%xmm1 \n"
650 "psrlq $0x8,%%xmm1 \n"
651 "por %%xmm1,%%xmm0 \n"
664 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
676 "movdqa 0x10(%0),%%xmm1 \n"
680 "pmaddubsw %%xmm4,%%xmm1 \n"
684 "phaddw %%xmm1,%%xmm0 \n"
701 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
713 "movdqu 0x10(%0),%%xmm1 \n"
717 "pmaddubsw %%xmm4,%%xmm1 \n"
721 "phaddw %%xmm1,%%xmm0 \n"
738 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
764 "movdqa 0x10(%0),%%xmm1 \n"
768 "pavgb 0x10(%0,%4,1),%%xmm1 \n"
773 "shufps $0x88,%%xmm1,%%xmm0 \n"
774 "shufps $0xdd,%%xmm1,%%xmm7 \n"
780 "movdqa %%xmm0,%%xmm1 \n"
784 "pmaddubsw %%xmm3,%%xmm1 \n"
787 "phaddw %%xmm6,%%xmm1 \n"
789 "psraw $0x8,%%xmm1 \n"
790 "packsswb %%xmm1,%%xmm0 \n"
804 , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
825 "movdqu 0x10(%0),%%xmm1 \n"
831 "pavgb %%xmm7,%%xmm1 \n"
838 "shufps $0x88,%%xmm1,%%xmm0 \n"
839 "shufps $0xdd,%%xmm1,%%xmm7 \n"
845 "movdqa %%xmm0,%%xmm1 \n"
849 "pmaddubsw %%xmm3,%%xmm1 \n"
852 "phaddw %%xmm6,%%xmm1 \n"
854 "psraw $0x8,%%xmm1 \n"
855 "packsswb %%xmm1,%%xmm0 \n"
869 , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
881 "movdqa 0x10(%0),%%xmm1 \n"
885 "pmaddubsw %%xmm4,%%xmm1 \n"
889 "phaddw %%xmm1,%%xmm0 \n"
906 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
918 "movdqu 0x10(%0),%%xmm1 \n"
922 "pmaddubsw %%xmm4,%%xmm1 \n"
926 "phaddw %%xmm1,%%xmm0 \n"
943 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
964 "movdqa 0x10(%0),%%xmm1 \n"
968 "pavgb 0x10(%0,%4,1),%%xmm1 \n"
973 "shufps $0x88,%%xmm1,%%xmm0 \n"
974 "shufps $0xdd,%%xmm1,%%xmm7 \n"
980 "movdqa %%xmm0,%%xmm1 \n"
984 "pmaddubsw %%xmm3,%%xmm1 \n"
987 "phaddw %%xmm6,%%xmm1 \n"
989 "psraw $0x8,%%xmm1 \n"
990 "packsswb %%xmm1,%%xmm0 \n"
1004 , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
1025 "movdqu 0x10(%0),%%xmm1 \n"
1031 "pavgb %%xmm7,%%xmm1 \n"
1038 "shufps $0x88,%%xmm1,%%xmm0 \n"
1039 "shufps $0xdd,%%xmm1,%%xmm7 \n"
1045 "movdqa %%xmm0,%%xmm1 \n"
1049 "pmaddubsw %%xmm3,%%xmm1 \n"
1052 "phaddw %%xmm6,%%xmm1 \n"
1054 "psraw $0x8,%%xmm1 \n"
1055 "packsswb %%xmm1,%%xmm0 \n"
1069 , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
1081 "movdqa 0x10(%0),%%xmm1 \n"
1085 "pmaddubsw %%xmm4,%%xmm1 \n"
1089 "phaddw %%xmm1,%%xmm0 \n"
1106 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1118 "movdqu 0x10(%0),%%xmm1 \n"
1122 "pmaddubsw %%xmm4,%%xmm1 \n"
1126 "phaddw %%xmm1,%%xmm0 \n"
1143 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1164 "movdqa 0x10(%0),%%xmm1 \n"
1168 "pavgb 0x10(%0,%4,1),%%xmm1 \n"
1173 "shufps $0x88,%%xmm1,%%xmm0 \n"
1174 "shufps $0xdd,%%xmm1,%%xmm7 \n"
1180 "movdqa %%xmm0,%%xmm1 \n"
1184 "pmaddubsw %%xmm3,%%xmm1 \n"
1187 "phaddw %%xmm6,%%xmm1 \n"
1189 "psraw $0x8,%%xmm1 \n"
1190 "packsswb %%xmm1,%%xmm0 \n"
1204 , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
1225 "movdqu 0x10(%0),%%xmm1 \n"
1231 "pavgb %%xmm7,%%xmm1 \n"
1238 "shufps $0x88,%%xmm1,%%xmm0 \n"
1239 "shufps $0xdd,%%xmm1,%%xmm7 \n"
1245 "movdqa %%xmm0,%%xmm1 \n"
1249 "pmaddubsw %%xmm3,%%xmm1 \n"
1252 "phaddw %%xmm6,%%xmm1 \n"
1254 "psraw $0x8,%%xmm1 \n"
1255 "packsswb %%xmm1,%%xmm0 \n"
1269 , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
1321 "movq (%[u_buf],%[v_buf],1),%%xmm1 \n" \
1323 "punpcklbw %%xmm1,%%xmm0 \n" \
1328 "movd (%[u_buf],%[v_buf],1),%%xmm1 \n" \
1330 "punpcklbw %%xmm1,%%xmm0 \n" \
1336 "movd (%[u_buf],%[v_buf],1),%%xmm1 \n" \
1338 "punpcklbw %%xmm1,%%xmm0 \n" \
1350 "movdqa %%xmm0,%%xmm1 \n" \
1353 "pmaddubsw 16(%[kYuvConstants]),%%xmm1 \n" \
1356 "psubw 64(%[kYuvConstants]),%%xmm1 \n" \
1364 "paddsw %%xmm3,%%xmm1 \n" \
1367 "psraw $0x6,%%xmm1 \n" \
1370 "packuswb %%xmm1,%%xmm1 \n" \
1375 "movdqa %%xmm0,%%xmm1 \n" \
1378 "pmaddubsw 144(%[kYuvConstants]),%%xmm1 \n" \
1381 "psubw 64(%[kYuvConstants]),%%xmm1 \n" \
1389 "paddsw %%xmm3,%%xmm1 \n" \
1392 "psraw $0x6,%%xmm1 \n" \
1395 "packuswb %%xmm1,%%xmm1 \n" \
1411 "punpcklbw %%xmm1,%%xmm0 \n"
1413 "movdqa %%xmm0,%%xmm1 \n"
1415 "punpckhwd %%xmm2,%%xmm1 \n"
1417 "movdqa %%xmm1,0x10(%[argb_buf]) \n"
1429 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1447 "punpcklbw %%xmm1,%%xmm0 \n"
1449 "movdqa %%xmm0,%%xmm1 \n"
1451 "punpckhwd %%xmm2,%%xmm1 \n"
1453 "movdqa %%xmm1,0x10(%[argb_buf]) \n"
1465 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1483 "punpcklbw %%xmm1,%%xmm0 \n"
1485 "movdqa %%xmm0,%%xmm1 \n"
1487 "punpckhwd %%xmm2,%%xmm1 \n"
1489 "movdqa %%xmm1,0x10(%[argb_buf]) \n"
1501 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1517 "punpcklbw %%xmm1,%%xmm0 \n"
1519 "movdqa %%xmm0,%%xmm1 \n"
1521 "punpckhwd %%xmm2,%%xmm1 \n"
1523 "movdqa %%xmm1,0x10(%[argb_buf]) \n"
1534 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1550 "punpcklbw %%xmm1,%%xmm0 \n"
1552 "movdqa %%xmm0,%%xmm1 \n"
1554 "punpckhwd %%xmm2,%%xmm1 \n"
1556 "movdqa %%xmm1,0x10(%[argb_buf]) \n"
1567 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1585 "punpcklbw %%xmm1,%%xmm0 \n"
1587 "movdqa %%xmm0,%%xmm1 \n"
1589 "punpckhwd %%xmm2,%%xmm1 \n"
1591 "movdqu %%xmm1,0x10(%[argb_buf]) \n"
1603 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1621 "punpcklbw %%xmm1,%%xmm0 \n"
1623 "movdqa %%xmm0,%%xmm1 \n"
1625 "punpckhwd %%xmm2,%%xmm1 \n"
1627 "movdqu %%xmm1,0x10(%[argb_buf]) \n"
1639 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1657 "punpcklbw %%xmm1,%%xmm0 \n"
1659 "movdqa %%xmm0,%%xmm1 \n"
1661 "punpckhwd %%xmm2,%%xmm1 \n"
1663 "movdqu %%xmm1,0x10(%[argb_buf]) \n"
1675 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1691 "punpcklbw %%xmm1,%%xmm0 \n"
1693 "movdqa %%xmm0,%%xmm1 \n"
1695 "punpckhwd %%xmm2,%%xmm1 \n"
1697 "movdqu %%xmm1,0x10(%[argb_buf]) \n"
1708 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1724 "punpcklbw %%xmm1,%%xmm0 \n"
1726 "movdqa %%xmm0,%%xmm1 \n"
1728 "punpckhwd %%xmm2,%%xmm1 \n"
1730 "movdqu %%xmm1,0x10(%[argb_buf]) \n"
1741 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1760 "punpcklbw %%xmm0,%%xmm1 \n"
1763 "punpcklwd %%xmm1,%%xmm5 \n"
1764 "punpckhwd %%xmm1,%%xmm0 \n"
1778 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1796 "punpcklbw %%xmm1,%%xmm2 \n"
1798 "movdqa %%xmm2,%%xmm1 \n"
1800 "punpckhwd %%xmm0,%%xmm1 \n"
1802 "movdqa %%xmm1,0x10(%[argb_buf]) \n"
1814 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1833 "punpcklbw %%xmm0,%%xmm1 \n"
1836 "punpcklwd %%xmm1,%%xmm5 \n"
1837 "punpckhwd %%xmm1,%%xmm0 \n"
1851 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1869 "punpcklbw %%xmm1,%%xmm2 \n"
1871 "movdqa %%xmm2,%%xmm1 \n"
1873 "punpckhwd %%xmm0,%%xmm1 \n"
1875 "movdqu %%xmm1,0x10(%[argb_buf]) \n"
1887 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1918 "movdqa %%xmm0,%%xmm1 \n"
1920 "punpckhwd %%xmm1,%%xmm1 \n"
1922 "por %%xmm4,%%xmm1 \n"
1924 "movdqa %%xmm1,16(%1) \n"
1935 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
1980 "movdqa %%xmm0,%%xmm1 \n"
1982 "psrlw $0x8,%%xmm1 \n"
1983 "por %%xmm1,%%xmm0 \n"
1997 , "xmm0", "xmm1"
2012 "movdqa %4,%%xmm1 \n"
2019 "pshufb %%xmm1,%%xmm0 \n"
2032 , "xmm0", "xmm1"
2078 "movdqa 0x10(%0),%%xmm1 \n"
2081 "movdqa %%xmm1,%%xmm3 \n"
2083 "pand %%xmm5,%%xmm1 \n"
2084 "packuswb %%xmm1,%%xmm0 \n"
2100 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
2113 "movdqa 0x10(%0),%%xmm1 \n"
2115 "movdqa %%xmm1,0x10(%0,%1) \n"
2125 , "xmm0", "xmm1"
2182 "movdqa 0x10(%0),%%xmm1 \n"
2185 "pand %%xmm5,%%xmm1 \n"
2186 "packuswb %%xmm1,%%xmm0 \n"
2197 , "xmm0", "xmm1", "xmm5"
2211 "movdqa 0x10(%0),%%xmm1 \n"
2216 "pavgb %%xmm3,%%xmm1 \n"
2218 "psrlw $0x8,%%xmm1 \n"
2219 "packuswb %%xmm1,%%xmm0 \n"
2220 "movdqa %%xmm0,%%xmm1 \n"
2223 "psrlw $0x8,%%xmm1 \n"
2224 "packuswb %%xmm1,%%xmm1 \n"
2226 "movq %%xmm1,(%1,%2) \n"
2237 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
2251 "movdqa 0x10(%0),%%xmm1 \n"
2254 "psrlw $0x8,%%xmm1 \n"
2255 "packuswb %%xmm1,%%xmm0 \n"
2256 "movdqa %%xmm0,%%xmm1 \n"
2259 "psrlw $0x8,%%xmm1 \n"
2260 "packuswb %%xmm1,%%xmm1 \n"
2262 "movq %%xmm1,(%1,%2) \n"
2273 , "xmm0", "xmm1", "xmm5"
2286 "movdqu 0x10(%0),%%xmm1 \n"
2289 "pand %%xmm5,%%xmm1 \n"
2290 "packuswb %%xmm1,%%xmm0 \n"
2301 , "xmm0", "xmm1", "xmm5"
2316 "movdqu 0x10(%0),%%xmm1 \n"
2321 "pavgb %%xmm3,%%xmm1 \n"
2323 "psrlw $0x8,%%xmm1 \n"
2324 "packuswb %%xmm1,%%xmm0 \n"
2325 "movdqa %%xmm0,%%xmm1 \n"
2328 "psrlw $0x8,%%xmm1 \n"
2329 "packuswb %%xmm1,%%xmm1 \n"
2331 "movq %%xmm1,(%1,%2) \n"
2342 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
2356 "movdqu 0x10(%0),%%xmm1 \n"
2359 "psrlw $0x8,%%xmm1 \n"
2360 "packuswb %%xmm1,%%xmm0 \n"
2361 "movdqa %%xmm0,%%xmm1 \n"
2364 "psrlw $0x8,%%xmm1 \n"
2365 "packuswb %%xmm1,%%xmm1 \n"
2367 "movq %%xmm1,(%1,%2) \n"
2378 , "xmm0", "xmm1", "xmm5"
2388 "movdqa 0x10(%0),%%xmm1 \n"
2391 "psrlw $0x8,%%xmm1 \n"
2392 "packuswb %%xmm1,%%xmm0 \n"
2403 , "xmm0", "xmm1"
2417 "movdqa 0x10(%0),%%xmm1 \n"
2422 "pavgb %%xmm3,%%xmm1 \n"
2424 "pand %%xmm5,%%xmm1 \n"
2425 "packuswb %%xmm1,%%xmm0 \n"
2426 "movdqa %%xmm0,%%xmm1 \n"
2429 "psrlw $0x8,%%xmm1 \n"
2430 "packuswb %%xmm1,%%xmm1 \n"
2432 "movq %%xmm1,(%1,%2) \n"
2443 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
2457 "movdqa 0x10(%0),%%xmm1 \n"
2460 "pand %%xmm5,%%xmm1 \n"
2461 "packuswb %%xmm1,%%xmm0 \n"
2462 "movdqa %%xmm0,%%xmm1 \n"
2465 "psrlw $0x8,%%xmm1 \n"
2466 "packuswb %%xmm1,%%xmm1 \n"
2468 "movq %%xmm1,(%1,%2) \n"
2479 , "xmm0", "xmm1", "xmm5"
2490 "movdqu 0x10(%0),%%xmm1 \n"
2493 "psrlw $0x8,%%xmm1 \n"
2494 "packuswb %%xmm1,%%xmm0 \n"
2505 , "xmm0", "xmm1"
2519 "movdqu 0x10(%0),%%xmm1 \n"
2524 "pavgb %%xmm3,%%xmm1 \n"
2526 "pand %%xmm5,%%xmm1 \n"
2527 "packuswb %%xmm1,%%xmm0 \n"
2528 "movdqa %%xmm0,%%xmm1 \n"
2531 "psrlw $0x8,%%xmm1 \n"
2532 "packuswb %%xmm1,%%xmm1 \n"
2534 "movq %%xmm1,(%1,%2) \n"
2545 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
2559 "movdqu 0x10(%0),%%xmm1 \n"
2562 "pand %%xmm5,%%xmm1 \n"
2563 "packuswb %%xmm1,%%xmm0 \n"
2564 "movdqa %%xmm0,%%xmm1 \n"
2567 "psrlw $0x8,%%xmm1 \n"
2568 "packuswb %%xmm1,%%xmm1 \n"
2570 "movq %%xmm1,(%1,%2) \n"
2581 , "xmm0", "xmm1", "xmm5"
2619 "movd (%1),%%xmm1 \n"
2621 "psrlw $0x8,%%xmm1 \n"
2623 "pmullw %%xmm3,%%xmm1 \n"
2626 "pand %%xmm5,%%xmm1 \n"
2627 "paddusb %%xmm1,%%xmm0 \n"
2651 "movdqu (%1),%%xmm1 \n"
2653 "psrlw $0x8,%%xmm1 \n"
2655 "pmullw %%xmm3,%%xmm1 \n"
2658 "pand %%xmm5,%%xmm1 \n"
2659 "paddusb %%xmm1,%%xmm0 \n"
2682 "movd (%1),%%xmm1 \n"
2684 "psrlw $0x8,%%xmm1 \n"
2686 "pmullw %%xmm3,%%xmm1 \n"
2689 "pand %%xmm5,%%xmm1 \n"
2690 "paddusb %%xmm1,%%xmm0 \n"
2703 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
2754 "movd (%1),%%xmm1 \n"
2756 "psrlw $0x8,%%xmm1 \n"
2758 "pmullw %%xmm3,%%xmm1 \n"
2761 "pand %%xmm5,%%xmm1 \n"
2762 "paddusb %%xmm1,%%xmm0 \n"
2788 "movdqa (%1),%%xmm1 \n"
2790 "psrlw $0x8,%%xmm1 \n"
2792 "pmullw %%xmm3,%%xmm1 \n"
2795 "pand %%xmm5,%%xmm1 \n"
2796 "paddusb %%xmm1,%%xmm0 \n"
2815 xmm1 \n"
2817 "psrlw $0x8,%%xmm1 \n"
2819 "pmullw %%xmm3,%%xmm1 \n"
2822 "pand %%xmm5,%%xmm1 \n"
2823 "paddusb %%xmm1,%%xmm0 \n"
2844 "movd (%1),%%xmm1 \n"
2846 "psrlw $0x8,%%xmm1 \n"
2848 "pmullw %%xmm3,%%xmm1 \n"
2851 "pand %%xmm5,%%xmm1 \n"
2852 "paddusb %%xmm1,%%xmm0 \n"
2865 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
2890 "movdqa (%0),%%xmm1 \n"
2891 "punpckhbw %%xmm1,%%xmm1 \n"
2892 "pshufhw $0xff,%%xmm1,%%xmm2 \n"
2894 "pmulhuw %%xmm2,%%xmm1 \n"
2898 "psrlw $0x8,%%xmm1 \n"
2899 "packuswb %%xmm1,%%xmm0 \n"
2912 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2942 "movdqa (%0),%%xmm1 \n"
2943 "punpcklbw %%xmm1,%%xmm1 \n"
2944 "pmulhuw %%xmm1,%%xmm0 \n"
2945 "movdqa (%0),%%xmm1 \n"
2946 "pshufb %%xmm5,%%xmm1 \n"
2949 "pmulhuw %%xmm2,%%xmm1 \n"
2953 "psrlw $0x8,%%xmm1 \n"
2954 "packuswb %%xmm1,%%xmm0 \n"
2967 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2997 "movdqa (%0),%%xmm1 \n"
2999 "punpckhbw %%xmm1,%%xmm1 \n"
3006 "pmulhuw %%xmm2,%%xmm1 \n"
3009 "packuswb %%xmm1,%%xmm0 \n"
3022 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
3044 "movdqa 0x10(%0),%%xmm1 \n"
3046 "pmaddubsw %%xmm4,%%xmm1 \n"
3047 "phaddw %%xmm1,%%xmm0 \n"
3059 "movdqa %%xmm0,%%xmm1 \n"
3061 "punpckhwd %%xmm3,%%xmm1 \n"
3064 "movdqa %%xmm1,0x10(%0,%1,1) \n"
3073 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
3114 "movdqa 0x10(%0),%%xmm1 \n"
3116 "pmaddubsw %%xmm3,%%xmm1 \n"
3117 "phaddw %%xmm1,%%xmm5 \n"
3122 "movdqa 0x10(%0),%%xmm1 \n"
3124 "pmaddubsw %%xmm4,%%xmm1 \n"
3125 "phaddw %%xmm1,%%xmm5 \n"
3129 "movdqa 0x10(%0),%%xmm1 \n"
3131 "psrld $0x18,%%xmm1 \n"
3132 "packuswb %%xmm1,%%xmm6 \n"
3135 "movdqa %%xmm0,%%xmm1 \n"
3137 "punpckhwd %%xmm5,%%xmm1 \n"
3140 "movdqa %%xmm1,0x10(%0) \n"
3150 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
3177 "movdqa 0x10(%0),%%xmm1 \n"
3179 "pmaddubsw %%xmm3,%%xmm1 \n"
3181 "phaddsw %%xmm1,%%xmm5 \n"
3188 "movdqa 0x10(%0),%%xmm1 \n"
3190 "pmaddubsw %%xmm4,%%xmm1 \n"
3191 "phaddsw %%xmm1,%%xmm5 \n"
3195 "movdqa 0x10(%0),%%xmm1 \n"
3197 "psrld $0x18,%%xmm1 \n"
3198 "packuswb %%xmm1,%%xmm6 \n"
3200 "movdqa %%xmm0,%%xmm1 \n"
3203 "punpckhwd %%xmm5,%%xmm1 \n"
3206 "movdqa %%xmm1,0x10(%0) \n"
3214 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
3245 "movdqa (%0),%%xmm1 \n"
3246 "punpckhbw %%xmm5,%%xmm1 \n"
3247 "pmulhuw %%xmm2,%%xmm1 \n"
3250 "pmullw %%xmm3,%%xmm1 \n"
3253 "paddw %%xmm4,%%xmm1 \n"
3254 "packuswb %%xmm1,%%xmm0 \n"
3267 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
3281 "pxor %%xmm1,%%xmm1 \n"
3293 "punpcklbw %%xmm1,%%xmm2 \n"
3295 "punpcklwd %%xmm1,%%xmm2 \n"
3296 "punpckhwd %%xmm1,%%xmm3 \n"
3297 "punpckhbw %%xmm1,%%xmm4 \n"
3299 "punpcklwd %%xmm1,%%xmm4 \n"
3300 "punpckhwd %%xmm1,%%xmm5 \n"
3330 "punpcklbw %%xmm1,%%xmm2 \n"
3331 "punpcklwd %%xmm1,%%xmm2 \n"
3348 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
3369 "movdqa 0x10(%0),%%xmm1 \n"
3373 "psubd 0x10(%0,%4,4),%%xmm1 \n"
3378 "psubd 0x10(%1),%%xmm1 \n"
3382 "paddd 0x10(%1,%4,4),%%xmm1 \n"
3387 "cvtdq2ps %%xmm1,%%xmm1 \n"
3389 "mulps %%xmm4,%%xmm1 \n"
3395 "cvtps2dq %%xmm1,%%xmm1 \n"
3398 "packssdw %%xmm1,%%xmm0 \n"
3437 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
3457 "movdqa %%xmm0,%%xmm1 \n"
3459 "punpckhbw %%xmm1,%%xmm1 \n"
3461 "pmulhuw %%xmm2,%%xmm1 \n"
3463 "psrlw $0x8,%%xmm1 \n"
3464 "packuswb %%xmm1,%%xmm0 \n"
3475 , "xmm0", "xmm1", "xmm2"
3517 "cvttps2dq %%xmm3,%%xmm1 \n"
3518 "packssdw %%xmm1,%%xmm0 \n"
3532 "movd (%0,%1,1),%%xmm1 \n"
3534 "punpckldq %%xmm6,%%xmm1 \n"
3536 "movq %%xmm1,(%2) \n"
3586 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
3614 "movdqa %%xmm0,%%xmm1 \n"
3616 "punpckhbw %%xmm2,%%xmm1 \n"
3618 "pmaddubsw %%xmm5,%%xmm1 \n"
3620 "psrlw $0x7,%%xmm1 \n"
3621 "packuswb %%xmm1,%%xmm0 \n"
3652 , "xmm0", "xmm1", "xmm2", "xmm5"