Lines Matching full:xmm2
228 "movdqa %%xmm3,%%xmm2 \n"
229 "palignr $0x8,%%xmm1,%%xmm2 \n"
230 "pshufb %%xmm4,%%xmm2 \n"
231 "por %%xmm5,%%xmm2 \n"
234 "movdqu %%xmm2," MEMACCESS2(0x20,1) " \n"
251 : "memory", "cc" , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
266 "movdqa %%xmm3,%%xmm2 \n"
267 "palignr $0x8,%%xmm1,%%xmm2 \n"
268 "pshufb %%xmm4,%%xmm2 \n"
269 "por %%xmm5,%%xmm2 \n"
272 "movdqu %%xmm2," MEMACCESS2(0x20,1) " \n"
289 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
302 "movdqu " MEMACCESS2(0x8,0) ",%%xmm2 \n"
306 "pshufb %%xmm5,%%xmm2 \n"
309 "movq %%xmm2," MEMACCESS2(0x10,1) " \n"
319 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
344 "movdqa %%xmm0,%%xmm2 \n"
346 "psllw $0xb,%%xmm2 \n"
348 "pmulhuw %%xmm5,%%xmm2 \n"
350 "por %%xmm2,%%xmm1 \n"
354 "movdqa %%xmm1,%%xmm2 \n"
356 "punpckhbw %%xmm0,%%xmm2 \n"
358 MEMOPMEM(movdqu,xmm2,0x10,1,0,2) // movdqu %%xmm2,0x10(%1,%0,2)
367 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
391 "movdqa %%xmm0,%%xmm2 \n"
393 "psllw $0xb,%%xmm2 \n"
395 "pmulhuw %%xmm5,%%xmm2 \n"
398 "por %%xmm2,%%xmm1 \n"
399 "movdqa %%xmm0,%%xmm2 \n"
401 "psraw $0x8,%%xmm2 \n"
403 "pand %%xmm7,%%xmm2 \n"
404 "por %%xmm2,%%xmm0 \n"
405 "movdqa %%xmm1,%%xmm2 \n"
407 "punpckhbw %%xmm0,%%xmm2 \n"
409 MEMOPMEM(movdqu,xmm2,0x10,1,0,2) // movdqu %%xmm2,0x10(%1,%0,2)
418 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
434 "movdqa %%xmm0,%%xmm2 \n"
436 xmm2 \n"
438 "movdqa %%xmm2,%%xmm3 \n"
442 "por %%xmm3,%%xmm2 \n"
444 "punpcklbw %%xmm2,%%xmm0 \n"
445 "punpckhbw %%xmm2,%%xmm1 \n"
456 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
467 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
472 "pshufb %%xmm6,%%xmm2 \n"
477 "movdqa %%xmm2,%%xmm5 \n"
482 "psrldq $0x8,%%xmm2 \n"
484 "por %%xmm3,%%xmm2 \n"
486 "movdqu %%xmm2," MEMACCESS2(0x20,1) " \n"
494 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
505 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
510 "pshufb %%xmm6,%%xmm2 \n"
515 "movdqa %%xmm2,%%xmm5 \n"
520 "psrldq $0x8,%%xmm2 \n"
522 "por %%xmm3,%%xmm2 \n"
524 "movdqu %%xmm2," MEMACCESS2(0x20,1) " \n"
532 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
549 "movdqa %%xmm0,%%xmm2 \n"
552 "psrld $0x5,%%xmm2 \n"
555 "pand %%xmm4,%%xmm2 \n"
557 "por %%xmm2,%%xmm1 \n"
568 :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
593 "movdqa %%xmm0,%%xmm2 \n"
596 "psrld $0x5,%%xmm2 \n"
599 "pand %%xmm4,%%xmm2 \n"
601 "por %%xmm2,%%xmm1 \n"
614 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
658 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
678 "movdqa %%xmm0,%%xmm2 \n"
682 "psrld $0x6,%%xmm2 \n"
686 "pand %%xmm5,%%xmm2 \n"
689 "por %%xmm3,%%xmm2 \n"
690 "por %%xmm2,%%xmm0 \n"
701 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
729 :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
744 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
748 "pmaddubsw %%xmm4,%%xmm2 \n"
752 "phaddw %%xmm3,%%xmm2 \n"
754 "psrlw $0x7,%%xmm2 \n"
755 "packuswb %%xmm2,%%xmm0 \n"
766 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
782 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
786 "pmaddubsw %%xmm4,%%xmm2 \n"
790 "phaddw %%xmm3,%%xmm2 \n"
792 "paddw %%xmm5,%%xmm2 \n"
794 "psrlw $0x7,%%xmm2 \n"
795 "packuswb %%xmm2,%%xmm0 \n"
805 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
851 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
893 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
914 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
916 "pavgb %%xmm7,%%xmm2 \n"
926 "movdqa %%xmm2,%%xmm7 \n"
927 "shufps $0x88,%%xmm6,%%xmm2 \n"
929 "pavgb %%xmm7,%%xmm2 \n"
931 "movdqa %%xmm2,%%xmm6 \n"
933 "pmaddubsw %%xmm4,%%xmm2 \n"
936 "phaddw %%xmm2,%%xmm0 \n"
956 "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
1021 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
1082 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
1103 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
1105 "pavgb %%xmm7,%%xmm2 \n"
1115 "movdqa %%xmm2,%%xmm7 \n"
1116 "shufps $0x88,%%xmm6,%%xmm2 \n"
1118 "pavgb %%xmm7,%%xmm2 \n"
1120 "movdqa %%xmm2,%%xmm6 \n"
1122 "pmaddubsw %%xmm4,%%xmm2 \n"
1125 "phaddw %%xmm2,%%xmm0 \n"
1146 "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
1163 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
1167 "pmaddubsw %%xmm4,%%xmm2 \n"
1170 "phaddw %%xmm6,%%xmm2 \n"
1172 "psraw $0x8,%%xmm2 \n"
1173 "packsswb %%xmm2,%%xmm0 \n"
1178 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
1182 "pmaddubsw %%xmm3,%%xmm2 \n"
1185 "phaddw %%xmm6,%%xmm2
1187 "psraw $0x8,%%xmm2 \n"
1188 "packsswb %%xmm2,%%xmm0 \n"
1203 "xmm0", "xmm1", "xmm2", "xmm6"
1216 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
1220 "pmaddubsw %%xmm4,%%xmm2 \n"
1224 "phaddw %%xmm3,%%xmm2 \n"
1226 "psrlw $0x7,%%xmm2 \n"
1227 "packuswb %%xmm2,%%xmm0 \n"
1238 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1257 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
1259 "pavgb %%xmm7,%%xmm2 \n"
1269 "movdqa %%xmm2,%%xmm7 \n"
1270 "shufps $0x88,%%xmm6,%%xmm2 \n"
1272 "pavgb %%xmm7,%%xmm2 \n"
1274 "movdqa %%xmm2,%%xmm6 \n"
1276 "pmaddubsw %%xmm4,%%xmm2 \n"
1279 "phaddw %%xmm2,%%xmm0 \n"
1299 "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
1311 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
1315 "pmaddubsw %%xmm4,%%xmm2 \n"
1319 "phaddw %%xmm3,%%xmm2 \n"
1321 "psrlw $0x7,%%xmm2 \n"
1322 "packuswb %%xmm2,%%xmm0 \n"
1333 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1345 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
1349 "pmaddubsw %%xmm4,%%xmm2 \n"
1353 "phaddw %%xmm3,%%xmm2 \n"
1355 "psrlw $0x7,%%xmm2 \n"
1356 "packuswb %%xmm2,%%xmm0 \n"
1367 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1386 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
1388 "pavgb %%xmm7,%%xmm2 \n"
1398 "movdqa %%xmm2,%%xmm7 \n"
1399 "shufps $0x88,%%xmm6,%%xmm2 \n"
1401 "pavgb %%xmm7,%%xmm2 \n"
1403 "movdqa %%xmm2,%%xmm6 \n"
1405 "pmaddubsw %%xmm4,%%xmm2 \n"
1408 "phaddw %%xmm2,%%xmm0 \n"
1428 "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
1447 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
1449 "pavgb %%xmm7,%%xmm2 \n"
1459 "movdqa %%xmm2,%%xmm7 \n"
1460 "shufps $0x88,%%xmm6,%%xmm2 \n"
1462 "pavgb %%xmm7,%%xmm2 \n"
1464 "movdqa %%xmm2,%%xmm6 \n"
1466 "pmaddubsw %%xmm4,%%xmm2 \n"
1469 "phaddw %%xmm2,%%xmm0 \n"
1489 "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
1595 "movdqa %%xmm0,%%xmm2 \n" \
1601 "pmaddubsw %%xmm9,%%xmm2 \n" \
1602 "psubw %%xmm2,%%xmm1 \n" \
1603 "movdqa %%xmm13,%%xmm2 \n" \
1605 "psubw %%xmm3,%%xmm2 \n" \
1609 "paddsw %%xmm4,%%xmm2 \n" \
1612 "psraw $0x6,%%xmm2 \n" \
1615 "packuswb %%xmm2,%%xmm2 \n"
1624 "movdqa %%xmm0,%%xmm2 \n" \
1630 "pmaddubsw " MEMACCESS2(32, [yuvconstants]) ",%%xmm2 \n" \
1631 "psubw %%xmm2,%%xmm1 \n" \
1632 "movdqa " MEMACCESS2(160, [yuvconstants]) ",%%xmm2 \n" \
1634 "psubw %%xmm3,%%xmm2 \n" \
1638 "paddsw %%xmm4,%%xmm2 \n" \
1641 "psraw $0x6,%%xmm2 \n" \
1644 "packuswb %%xmm2,%%xmm2 \n"
1651 "punpcklbw %%xmm5,%%xmm2 \n" \
1653 "punpcklwd %%xmm2,%%xmm0 \n" \
1654 "punpckhwd %%xmm2,%%xmm1 \n" \
1662 "punpcklbw %%xmm2,%%xmm1 \n" \
1695 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1715 "punpcklbw %%xmm2,%%xmm2 \n"
1717 "punpcklwd %%xmm2,%%xmm0 \n"
1718 "punpckhwd %%xmm2,%%xmm1 \n"
1740 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
1768 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1802 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1838 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1864 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1890 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1915 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1940 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
1968 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2161 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2194 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2227 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2265 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2308 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2338 "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2369 "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2399 "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2429 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2438 "movd %%eax,%%xmm2 \n"
2439 "pshufd $0x0,%%xmm2,%%xmm2 \n"
2451 "pmulhuw %%xmm2,%%xmm0 \n"
2474 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
2485 "vmovd %%eax,%%xmm2 \n"
2486 "vbroadcastss %%xmm2,%%ymm2 \n"
2521 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
2696 "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
2713 "movdqa %%xmm0,%%xmm2 \n"
2718 "psrlw $0x8,%%xmm2 \n"
2720 "packuswb %%xmm3,%%xmm2 \n"
2722 MEMOPMEM(movdqu,xmm2,0x00,1,2,1) // movdqu %%xmm2,(%1,%2)
2732 "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
2763 "xmm0", "xmm1", "xmm2"
2778 "movdqa %%xmm0,%%xmm2 \n"
2780 "punpckhbw %%xmm1,%%xmm2 \n"
2782 "movdqu %%xmm2," MEMACCESS2(0x10,2) " \n"
2792 "xmm0", "xmm1", "xmm2"
2884 "movdqu " MEMACCESS(0) ",%%xmm2 \n"
2889 "pand %%xmm0,%%xmm2 \n"
2893 "por %%xmm4,%%xmm2 \n"
2895 "movdqu %%xmm2," MEMACCESS(1) " \n"
2905 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
2934 , "xmm0", "xmm1", "xmm2"
2976 "movq " MEMACCESS(0) ",%%xmm2 \n"
2978 "punpcklbw %%xmm2,%%xmm2 \n"
2979 "punpckhwd %%xmm2,%%xmm3 \n"
2980 "punpcklwd %%xmm2,%%xmm2 \n"
2983 "pand %%xmm0,%%xmm2 \n"
2987 "por %%xmm4,%%xmm2 \n"
2989 "movdqu %%xmm2," MEMACCESS(1) " \n"
2999 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
3030 , "xmm0", "xmm1", "xmm2"
3104 MEMOPREG(movdqu,0x00,0,4,1,xmm2) // movdqu (%0,%4,1),%%xmm2
3107 "pavgb %%xmm2,%%xmm0 \n"
3128 "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
3199 MEMOPREG(movdqu,0x00,0,4,1,xmm2) // movdqu (%0,%4,1),%%xmm2
3202 "pavgb %%xmm2,%%xmm0 \n"
3223 "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
3497 "movdqu " MEMACCESS(1) ",%%xmm2 \n"
3499 "pand %%xmm6,%%xmm2 \n"
3501 "pmullw %%xmm3,%%xmm2 \n"
3507 "psrlw $0x8,%%xmm2 \n"
3508 "paddusb %%xmm2,%%xmm0 \n"
3526 "movd " MEMACCESS(1) ",%%xmm2 \n"
3528 "pand %%xmm6,%%xmm2 \n"
3530 "pmullw %%xmm3,%%xmm2 \n"
3536 "psrlw $0x8,%%xmm2 \n"
3537 "paddusb %%xmm2,%%xmm0 \n"
3551 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
3584 "movq (%1,%2,1),%%xmm2 \n"
3585 "punpcklbw %%xmm2,%%xmm1 \n"
3600 :: "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7"
3658 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
3690 "movdqu " MEMACCESS(0) ",%%xmm2 \n"
3691 "punpckhbw %%xmm2,%%xmm2 \n"
3692 "pmulhuw %%xmm2,%%xmm1 \n"
3693 "movdqu " MEMACCESS(0) ",%%xmm2 \n"
3695 "pand %%xmm3,%%xmm2 \n"
3699 "por %%xmm2,%%xmm0 \n"
3710 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
3753 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
3770 MEMOPREG(movd,0x00,4,3,4,xmm2) // movd 0x0(%4,%3,4),%%xmm2
3773 "pshuflw $0x40,%%xmm2,%%xmm2 \n"
3775 "movlhps %%xmm3,%%xmm2 \n"
3776 "pmulhuw %%xmm2,%%xmm0 \n"
3780 MEMOPREG(movd,0x00,4,3,4,xmm2) // movd 0x0(%4,%3,4),%%xmm2
3783 "pshuflw $0x40,%%xmm2,%%xmm2 \n"
3785 "movlhps %%xmm3,%%xmm2 \n"
3786 "pmulhuw %%xmm2,%%xmm1 \n"
3799 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
3827 MEMOPREG(vmovd,0x00,4,3,4,xmm2) // vmovd 0x0(%4,%3,4),%%xmm2
3831 "vpunpckldq %%xmm3,%%xmm2,%%xmm7 \n"
3837 MEMOPREG(vmovd,0x00,4,3,4,xmm2) // vmovd 0x0(%4,%3,4),%%xmm2
3840 "vpunpckldq %%xmm3,%%xmm2,%%xmm2 \n"
3842 "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0 \n"
3868 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
3891 "movdqu " MEMACCESS(0) ",%%xmm2 \n"
3894 "psrld $0x18,%%xmm2 \n"
3896 "packuswb %%xmm3,%%xmm2 \n"
3897 "packuswb %%xmm2,%%xmm2 \n"
3900 "punpcklbw %%xmm2,%%xmm3 \n"
3915 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
3940 "movdqa %2,%%xmm2 \n"
3949 "pmaddubsw %%xmm2,%%xmm0 \n"
3950 "pmaddubsw %%xmm2,%%xmm6 \n"
3990 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
4002 "pshufd $0x00,%%xmm5,%%xmm2 \n"
4012 "pmaddubsw %%xmm2,%%xmm0 \n"
4013 "pmaddubsw %%xmm2,%%xmm7 \n"
4054 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
4064 "movd %2,%%xmm2 \n"
4067 "pshuflw $0x40,%%xmm2,%%xmm2 \n"
4068 "pshufd $0x44,%%xmm2,%%xmm2 \n"
4082 "pmulhuw %%xmm2,%%xmm0 \n"
4085 "pmulhuw %%xmm2,%%xmm1 \n"
4104 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
4114 "movd %3,%%xmm2 \n"
4115 "punpcklbw %%xmm2,%%xmm2 \n"
4116 "punpcklqdq %%xmm2,%%xmm2 \n"
4126 "pmulhuw %%xmm2,%%xmm0 \n"
4127 "pmulhuw %%xmm2,%%xmm1 \n"
4140 , "xmm0", "xmm1", "xmm2"
4157 "movdqu " MEMACCESS(1) ",%%xmm2 \n"
4160 "movdqu %%xmm2,%%xmm3 \n"
4163 "punpcklbw %%xmm5,%%xmm2 \n"
4165 "pmulhuw %%xmm2,%%xmm0 \n"
4178 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
4216 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
4356 MEMOPREG(movq,0x02,0,1,1,xmm2) // movq 0x2(%0,%1,1),%%xmm2
4358 "punpcklbw %%xmm5,%%xmm2 \n"
4359 "psubw %%xmm2,%%xmm1 \n"
4360 MEMOPREG(movq,0x00,0,2,1,xmm2) // movq (%0,%2,1),%%xmm2
4362 "punpcklbw %%xmm5,%%xmm2 \n"
4364 "psubw %%xmm3,%%xmm2 \n"
4365 "paddw %%xmm2,%%xmm0 \n"
4383 "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
4409 MEMOPREG(movq,0x01,0,1,1,xmm2) // movq 0x1(%0,%1,1),%%xmm2
4411 "punpcklbw %%xmm5,%%xmm2 \n"
4412 "psubw %%xmm2,%%xmm1 \n"
4413 "movq " MEMACCESS2(0x2,0) ",%%xmm2 \n"
4415 "punpcklbw %%xmm5,%%xmm2 \n"
4417 "psubw %%xmm3,%%xmm2 \n"
4418 "paddw %%xmm2,%%xmm0 \n"
4435 "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
4460 "movdqa %%xmm0,%%xmm2 \n"
4461 "punpcklbw %%xmm0,%%xmm2 \n"
4463 "movdqa %%xmm2,%%xmm1 \n"
4464 "punpcklwd %%xmm2,%%xmm1 \n"
4465 "punpckhwd %%xmm2,%%xmm2 \n"
4467 "por %%xmm5,%%xmm2 \n"
4474 "movdqu %%xmm2," MEMACCESS2(0x10,2) " \n"
4486 "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
4540 "movdqa %%xmm0,%%xmm2 \n"
4541 "paddusb %%xmm1,%%xmm2 \n"
4546 "punpcklbw %%xmm2,%%xmm4 \n"
4547 "punpckhbw %%xmm2,%%xmm1 \n"
4567 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
4588 "movdqu " MEMACCESS(0) ",%%xmm2 \n"
4590 "movdqa %%xmm2,%%xmm4 \n"
4591 "punpcklbw %%xmm1,%%xmm2 \n"
4592 "movdqa %%xmm2,%%xmm3 \n"
4593 "punpcklwd %%xmm1,%%xmm2 \n"
4599 "paddd %%xmm2,%%xmm0 \n"
4600 "movdqu " MEMACCESS(2) ",%%xmm2 \n"
4601 "paddd %%xmm0,%%xmm2 \n"
4612 "movdqu %%xmm2," MEMACCESS(1) " \n"
4627 "movd " MEMACCESS(0) ",%%xmm2 \n"
4629 "punpcklbw %%xmm1,%%xmm2 \n"
4630 "punpcklwd %%xmm1,%%xmm2 \n"
4631 "paddd %%xmm2,%%xmm0 \n"
4632 "movdqu " MEMACCESS(2) ",%%xmm2 \n"
4634 "paddd %%xmm0,%%xmm2 \n"
4635 "movdqu %%xmm2," MEMACCESS(1) " \n"
4647 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
4680 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
4684 MEMOPREG(psubd,0x20,0,4,4,xmm2) // psubd 0x20(%0,%4,4),%%xmm2
4689 "psubd " MEMACCESS2(0x20,1) ",%%xmm2 \n"
4693 MEMOPREG(paddd,0x20,1,4,4,xmm2) // paddd 0x20(%1,%4,4),%%xmm2
4697 "packssdw %%xmm3,%%xmm2 \n"
4699 "pmulhuw %%xmm5,%%xmm2 \n"
4700 "packuswb %%xmm2,%%xmm0 \n"
4712 "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n"
4716 MEMOPREG(psubd,0x20,0,4,4,xmm2) // psubd 0x20(%0,%4,4),%%xmm2
4721 "psubd " MEMACCESS2(0x20,1) ",%%xmm2 \n"
4725 MEMOPREG(paddd,0x20,1,4,4,xmm2) // paddd 0x20(%1,%4,4),%%xmm2
4732 "cvtdq2ps %%xmm2,%%xmm2 \n"
4734 "mulps %%xmm4,%%xmm2 \n"
4738 "cvtps2dq %%xmm2,%%xmm2 \n"
4741 "packssdw %%xmm3,%%xmm2 \n"
4742 "packuswb %%xmm2,%%xmm0 \n"
4778 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
4791 "movq " MEMACCESS(3) ",%%xmm2 \n"
4801 "movdqa %%xmm2,%%xmm0 \n"
4803 "movlhps %%xmm0,%%xmm2 \n"
4806 "movdqa %%xmm2,%%xmm3 \n"
4813 "cvttps2dq %%xmm2,%%xmm0 \n" // x, y float to int first 2
4824 "addps %%xmm4,%%xmm2 \n"
4845 "cvttps2dq %%xmm2,%%xmm0 \n"
4848 "addps %%xmm7,%%xmm2 \n"
4864 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
4896 MEMOPREG(movdqu,0x00,1,4,1,xmm2)
4898 "punpcklbw %%xmm2,%%xmm0 \n"
4899 "punpckhbw %%xmm2,%%xmm1 \n"
4902 "movdqa %%xmm5,%%xmm2 \n"
4904 "pmaddubsw %%xmm0,%%xmm2 \n"
4906 "paddw %%xmm4,%%xmm2 \n"
4908 "psrlw $0x8,%%xmm2 \n"
4910 "packuswb %%xmm3,%%xmm2 \n"
4911 MEMOPMEM(movdqu,xmm2,0x00,1,0,1)
4945 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
5021 "xmm0", "xmm1", "xmm2", "xmm4", "xmm5"
5212 "movq " MEMACCESS(1) ",%%xmm2 \n"
5215 "punpcklbw %%xmm3,%%xmm2 \n"
5219 "punpcklbw %%xmm2,%%xmm0 \n"
5220 "punpckhbw %%xmm2,%%xmm1 \n"
5233 "xmm0", "xmm1", "xmm2", "xmm3"
5247 "movq " MEMACCESS(1) ",%%xmm2 \n"
5250 "punpcklbw %%xmm3,%%xmm2 \n"
5252 "movdqa %%xmm2,%%xmm1 \n"
5255 "punpckhbw %%xmm0,%%xmm2 \n"
5257 "movdqu %%xmm2," MEMACCESS2(0x10,3) " \n"
5268 "xmm0", "xmm1", "xmm2", "xmm3"
5297 "movdqa %%xmm1,%%xmm2 \n"
5299 "mulps %%xmm1,%%xmm2 \n"
5301 "mulps %%xmm2,%%xmm1 \n"
5303 "mulps " MEMACCESS2(0x20,3) ",%%xmm2 \n"
5307 "addps %%xmm2,%%xmm0 \n"
5324 , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
5364 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
