Lines Matching refs:m1
118 mova m1, [aq+mmsize]
124 mova [dstq+strideq*0+16], m1
126 mova [dstq+strideq*1+16], m1
128 mova [dstq+strideq*2+16], m1
130 mova [dstq+stride3q +16], m1
140 mova m1, [aq+mmsize*1]
147 mova [dstq+strideq*0+16], m1
151 mova [dstq+strideq*1+16], m1
165 pshufw m1, m3, q2222
169 mova [dstq+strideq*1], m1
181 pshufd m1, m3, q2222
183 mova [dstq+strideq*1], m1
185 pshufd m1, m3, q0000
187 mova [dstq+stride3q ], m1
191 pshufd m1, m2, q2222
193 mova [dstq+strideq*1], m1
195 pshufd m1, m2, q0000
197 mova [dstq+stride3q ], m1
208 pshufd m1, m3, q2222
213 mova [dstq+strideq*1+ 0], m1
214 mova [dstq+strideq*1+16], m1
232 pshufd m1, m3, q2222
239 mova [dstq+strideq*1+ 0], m1
240 mova [dstq+strideq*1+16], m1
241 mova [dstq+strideq*1+32], m1
242 mova [dstq+strideq*1+48], m1
263 pshufw m1, m0, q3232
265 paddd m0, m1
281 pshufd m1, m0, q3232
282 paddd m0, m1
283 pshufd m1, m0, q1111
285 paddd m0, m1
310 pshufd m1, m0, q3232
311 paddd m0, m1
312 pshufd m1, m0, q1111
314 paddd m0, m1
346 pshufd m1, m0, q3232
347 paddd m0, m1
348 pshufd m1, m0, q1111
350 paddd m0, m1
375 pshufw m1, m0, q3232
377 paddd m0, m1
392 pshufd m1, m0, q3232
393 paddd m0, m1
394 pshufd m1, m0, q1111
396 paddd m0, m1
419 pshufd m1, m0, q3232
420 paddd m0, m1
421 pshufd m1, m0, q1111
423 paddd m0, m1
450 pshufd m1, m0, q3232
451 paddd m0, m1
452 pshufd m1, m0, q1111
454 paddd m0, m1
488 pshufw m1, m3, q2222
492 paddw m1, m4
497 pmaxsw m1, m4
501 pminsw m1, m5
505 mova [dstq+strideq*1], m1
531 pshufd m1, m3, q2222
535 paddw m1, m5
539 pmaxsw m1, m6
543 pminsw m1, m4
547 mova [dstq+strideq*1], m1
580 paddw m1, m3, m4
584 pmaxsw m1, m6
588 pminsw m1, m7
592 mova [dstq+strideq*1+ 0], m1
607 pxor m1, m1
615 mova [rsp+16], m1
638 paddw m1, m3, m5
642 pmaxsw m1, reg_min
646 pminsw m1, reg_max
650 mova [dstq+strideq*0+16], m1
704 movu m1, [aq] ; abcdefgh
705 pshufhw m0, m1, q3310 ; abcdefhh
706 SHIFT_RIGHT m1, m1 ; bcdefghh
707 psrldq m2, m1, 2 ; cdefghh.
709 pshufd m1, m0, q3321 ; DEFGh...
711 movh [dstq+strideq*2], m1
714 psrldq m1, 2 ; EFGh....
716 movh [dstq+strideq*2], m1
725 SHIFT_RIGHTx2 m1, m2, m0, m4 ; bcdefghh/cdefghhh
727 shufps m1, m0, m2, q3332 ; FGHhhhhh
728 shufps m3, m0, m1, q2121 ; DEFGHhhh
733 mova [dstq+strideq*4], m1
735 pshuflw m1, m1, q3321 ; GHhhhhhh
738 mova [dstq+stride5q ], m1
740 pshuflw m1, m1, q3321 ; Hhhhhhhh
742 mova [dstq+strideq*4], m1
743 pshuflw m1, m1, q3321 ; hhhhhhhh
745 mova [dstq+stride5q ], m1
752 PALIGNR m1, m3, m0, 2, m4 ; bcdefghi
758 SHIFT_RIGHTx2 m2, m1, m3, m4 ; jklmnopp/klmnoppp
766 mova [dstq+strideq*0+16], m1
767 mova [dstq+strideq*8+ 0], m1
771 vpalignr m0, m1, m0, 2
773 PALIGNR m3, m1, m0, 2, m4
776 SHIFT_RIGHT m1, m1, m4
784 mova m1, [aq+mmsize*1] ; ijklmnop
787 PALIGNR m4, m1, m0, 2, m6
788 PALIGNR m5, m1, m0, 4, m6
790 PALIGNR m4, m2, m1, 2, m6
791 PALIGNR m5, m2, m1, 4, m6
809 mova [dstq+stride8q*0+16], m1
812 mova [dstq+stride8q*1+ 0], m1
826 vpalignr m0, m1, m0, 2
827 vpalignr m1, m2, m1, 2
830 PALIGNR m5, m1, m0, 2, m6
832 PALIGNR m5, m2, m1, 2, m6
833 mova m1, m5
856 vperm2i128 m2, m0, m1, q0201 ; ijklmnoppppppppp
860 vperm2i128 m2, m0, m1, q0201 ; JKLMNOPppppppppp
892 mova m1, [aq+mmsize*1+ 0] ; qrstuvwxyz012345
894 vperm2i128 m5, m0, m1, q0201 ; ijklmnopqrstuvwx
898 vperm2i128 m5, m1, m4, q0201 ; yz01234555555555
899 vpalignr m2, m5, m1, 2 ; rstuvwxyz0123455
900 vpalignr m3, m5, m1, 4 ; stuvwxyz01234555
902 vperm2i128 m2, m1, m4, q0201 ; Z......555555555
903 vperm2i128 m5, m0, m1, q0201 ; JKLMNOPQRSTUVWXY
910 mova [dstq+strideq*0 +32], m1
912 vpalignr m4, m2, m1, 2
916 vpalignr m4, m2, m1, 4
920 vpalignr m4, m2, m1, 6
925 vpalignr m4, m2, m1, 8
929 vpalignr m4, m2, m1, 10
933 vpalignr m4, m2, m1, 12
937 vpalignr m4, m2, m1, 14
941 vpalignr m4, m2, m1, 16
945 mova m1, m4
956 movd m1, [aq+6] ; d.......
957 PALIGNR m1, m0, 2, m2 ; xyz*abcd
958 psrldq m2, m1, 2 ; yz*abcd.
974 movu m1, [aq-2] ; *abcdefg
978 PALIGNR m1, m0, 2, m4 ; tuvwxyz*
979 PALIGNR m2, m1, 2, m4 ; uvwxyz*a
988 PALIGNR m1, m3, m2, 2, m0
990 movhps [dstq +strideq*2+0], m1
992 mova [dst4q+strideq*2+0], m1
993 PALIGNR m2, m3, m1, 2, m0
998 PALIGNR m1, m3, m2, 2, m0
1000 movhps [dstq +strideq*0+0], m1
1002 mova [dst4q+strideq*0+0], m1
1007 mova m1, [lq+mmsize] ; stuvwxyz
1016 PALIGNR m2, m1, 2, m3 ; tuvwxyz*
1019 PALIGNR m1, m0, 2, m3 ; lmnopqrs
1020 PALIGNR m2, m1, 2, m3 ; mnopqrst
1037 PALIGNR m0, m4, m2, 2, m1
1039 PALIGNR m0, m5, m4, 2, m1
1041 PALIGNR m0, m6, m5, 2, m1
1052 movu m1, [aq+mmsize*3-2] ; a[23-30]
1055 mova m1, [aq+mmsize*2] ; a[16-23]
1057 PALIGNR m0, m1, 2, m4 ; a[17-24]
1061 PALIGNR m1, m3, 2, m5 ; a[9-16]
1080 mova m1, [lq+mmsize*1] ; l[8-15]
1081 PALIGNR m7, m1, 2, m0 ; l[9-16]
1085 PALIGNR m1, m3, 2, m0 ; l[1-8]
1086 PALIGNR m7, m1, 2, m0 ; l[2-9]
1112 mova [dst8q+stride8q*0+32], m1
1117 mova [dst8q+stride8q*1+48], m1
1131 vpalignr m3, m1, m3, 2
1132 vpalignr m1, m0, m1, 2
1147 PALIGNR m2, m1, m3, 2, m0
1153 PALIGNR m2, m0, m1, 2, m3
1154 mova m1, m2
1157 PALIGNR m1, m2, m0, 2, m3
1158 mova m0, m1
1177 movu m1, [aq-2] ; *abcdefghijklmno
1181 vperm2i128 m3, m0, m1, q0201 ; stuvwxyz*abcdefg
1186 vperm2i128 m5, m0, m1, q0201 ; TUVWXYZ#ABCDEFGH
1193 vpalignr m4, m1, m5, 2
1197 vpalignr m4, m1, m5, 4
1203 vpalignr m4, m1, m5, 6
1207 vpalignr m4, m1, m5, 8
1211 vpalignr m4, m1, m5, 10
1215 vpalignr m4, m1, m5, 12
1219 vpalignr m4, m1, m5, 14
1229 vperm2i128 m2, m0, m1, q0201 ; ijklmnoppppppppp
1231 vperm2i128 m4, m3, m1, q0201 ; jklmnopppppppppp
1233 vperm2i128 m6, m5, m1, q0201 ; klmnoppppppppppp
1246 vpalignr m1, m6, m5, 2
1248 mova [dstq+stride3q*1], m1 ; 3 CDEFGHIJKLMNOPPP
1250 vpalignr m1, m6, m5, 4
1252 mova [dstq+stride5q*1], m1 ; 5 DEFGHIJKLMNOPPPP
1254 vpalignr m1, m6, m5, 6
1256 mova [dst4q+stride3q*1], m1 ; 7 EFGHIJKLMNOPPPPP
1258 vpalignr m1, m6, m5, 8
1260 mova [dst4q+stride5q*1], m1 ; 9 FGHIJKLMNOPPPPPP
1268 vpalignr m1, m6, m5, 10
1269 mova [dst4q+strideq*8], m1 ; 11 GHIJKLMNOPPPPPPP
1270 vpalignr m1, m6, m5, 12
1271 mova [dst4q+stride5q*2], m1 ; 13 HIJKLMNOPPPPPPPP
1272 vpalignr m1, m6, m5, 14
1273 mova [dst4q+stride3q*4], m1 ; 15 IJKLMNOPPPPPPPPP
1278 mova m1, [lq] ; klmnopqrstuvwxyz
1279 vperm2i128 m2, m1, m0, q0201 ; stuvwxyz*abcdefg
1280 vpalignr m3, m2, m1, 2 ; lmnopqrstuvwxyz*
1281 vpalignr m4, m2, m1, 4 ; mnopqrstuvwxyz*a
1283 pavgw m3, m1 ; klmnopqrstuvwxyz
1284 mova m1, [aq] ; abcdefghijklmnop
1288 vpunpckhwd m1, m3, m4 ; oPpQqRrSwXxYyZz#
1289 vperm2i128 m3, m1, m0, q0002 ; kLlMmNnOoPpQqRrS
1290 vperm2i128 m4, m0, m1, q0301 ; sTtUuVvWwXxYyZz#
1292 vperm2i128 m1, m3, m4, q0201 ; oPpQqRrSsTtUuVvW
1299 mova [dst5q+stride3q*2], m1 ; 11 oPpQqRrSsTtUuVvW
1302 vpalignr m5, m4, m1, 4
1308 vpalignr m5, m4, m1, 8
1314 vpalignr m5, m1, m3, 12
1315 vpalignr m6, m4, m1, 12
1323 vpalignr m5, m1, m3, 4
1326 vpalignr m5, m1, m3, 8
1333 mova m1, [lq+mmsize*1+0] ; l[16-31]
1337 vperm2i128 m5, m0, m1, q0201 ; lmnopqrstuvwxyz0
1341 vperm2i128 m7, m1, m2, q0201 ; stuvwxyz*abcdefg
1342 vpalignr m5, m7, m1, 2 ; lmnopqrstuvwxyz*
1343 vpalignr m6, m7, m1, 4 ; mnopqrstuvwxyz*a
1352 vperm2i128 m4, m1, m2, q0201 ; TUVWXYZ#ABCDEFGH
1353 vperm2i128 m5, m0, m1, q0201 ; L[7-15]L[16-23]
1365 mova [dst24q+stride7q+32], m1
1366 mova [dst8q+stride7q+0], m1
1368 vpalignr m6, m4, m1, 2
1375 vpalignr m6, m4, m1, 4
1382 vpalignr m6, m4, m1, 6
1389 vpalignr m6, m4, m1, 8
1396 vpalignr m6, m4, m1, 10
1403 vpalignr m6, m4, m1, 12
1410 vpalignr m6, m4, m1, 14
1418 mova m5, m1
1419 mova m1, m4
1437 psrldq m1, m0, 2 ; bcdefgh.
1440 pavgw m1, m0 ; ABCDEFG.
1444 movh [dstq+strideq*0], m1
1446 psrldq m1, 2
1448 movh [dstq+strideq*2], m1
1458 SHIFT_RIGHTx2 m1, m2, m0, m3 ; bcdefghh/cdefghhh
1460 pavgw m1, m0 ; ABCDEFGh
1464 mova [dstq+strideq*0], m1
1466 SHIFT_RIGHT m1, m1, m3
1468 mova [dstq+strideq*2], m1
1471 SHIFT_RIGHT m1, m1, m3
1473 mova [dstq+strideq*0], m1
1475 SHIFT_RIGHT m1, m1, m3
1477 mova [dstq+strideq*2], m1
1484 mova m1, [aq+mmsize]
1485 PALIGNR m2, m1, m0, 2, m3
1486 PALIGNR m3, m1, m0, 4, m4
1492 SHIFT_RIGHTx2 m5, m0, m1, m4
1494 pavgw m1, m5
1500 mova [dstq+strideq*0+16], m1
1505 vpalignr m2, m1, m2, 2
1508 PALIGNR m5, m1, m2, 2, m4
1513 SHIFT_RIGHT m1, m1, m4
1522 mova m1, [aq+mmsize*1]
1524 PALIGNR m6, m1, m0, 2, m5
1525 PALIGNR m7, m1, m0, 4, m5
1529 PALIGNR m4, m2, m1, 2, m0
1530 PALIGNR m5, m2, m1, 4, m0
1532 pavgw m4, m1
1534 PALIGNR m1, m0, m2, 2, m6
1537 pavgw m2, m1
1541 SHIFT_RIGHTx2 m6, m1, m0, reg_shuf
1568 mova [dstq+strideq*1+48], m1
1577 mova [dstq+stride17q+32], m1
1588 vpalignr m3, m1, m3, 2
1594 PALIGNR m3, m4, m6, 2, m1
1596 PALIGNR m3, m2, m4, 2, m1
1598 PALIGNR m3, m0, m2, 2, m1
1600 PALIGNR m3, m5, m7, 2, m1
1610 PALIGNR m6, m1, m3, 2, m7
1616 SHIFT_RIGHT m1, m1, reg_shuf
1649 movhps m1, [lq]
1650 PALIGNR m0, m1, 10, m2 ; xyz*abcd
1651 pslldq m1, m0, 2 ; .xyz*abc
1654 pavgw m1, m0 ; ....#ABC
1658 movhps [dstq+strideq*0], m1
1660 shufps m0, m2, m1, q3210
1673 movu m1, [aq-2] ; *abcdefg
1676 PALIGNR m3, m1, m2, 14, m4 ; z*abcdef
1678 pavgw m0, m1
1679 PALIGNR m1, m2, 2, m4 ; tuvwxyz*
1687 PALIGNR m0, m4, 14, m1
1689 PALIGNR m3, m4, 14, m1
1694 PALIGNR m0, m4, 14, m1
1696 PALIGNR m3, m4, 14, m1
1700 PALIGNR m0, m4, 14, m1
1708 movu m1, [aq-2] ; *abcdefg
1713 PALIGNR m0, m1, m5, 14, m6 ; z*abcdef
1718 pavgw m3, m1
1719 PALIGNR m1, m5, 2, m7 ; tuvwxyz*
1726 psrld m4, m1, 16
1728 pand m1, [pd_65535]
1731 packssdw m5, m1
1753 movu m1, [aq+mmsize*1-2] ; a[7-14]
1770 pavgw m1, m4
1817 mova [dstq+strideq*0+16], m1
1828 mova [dstq+stride16q+32], m1
1838 PALIGNR m2, m1, 14, m5
1839 PALIGNR m1, m4, 14, m5
1867 mova [dstq+strideq*0+48], m1
1873 PALIGNR m1, m2, 14, m6
1901 psrldq m1, m0, 2 ; bcddddd.
1904 pavgw m1, m0 ; abcddddd
1906 PALIGNR m2, m1, 4, m0 ; bCcDdddd
1910 movh [dstq+strideq*0], m1 ; aBbC
1912 movhps [dstq+strideq*2], m1 ; cDdd
1921 SHIFT_RIGHTx2 m1, m2, m0, m3
1923 pavgw m1, m0
1925 shufps m0, m1, m2, q1032
1930 mova [dstq+strideq *0], m1
1936 vpalignr m1, m2, m1, 4
1938 PALIGNR m0, m2, m1, 4, m3
1939 mova m1, m0
1942 shufps m0, m1, m2, q1032
1944 mova [dstq+strideq *0], m1
1953 movu m1, [lq+2]
1956 pavgw m1, m0
1971 mova [dstq+strideq *0+ 0], m1
1981 vpalignr m1, m2, m1, 4
1986 PALIGNR m5, m2, m1, 4, m6
1987 mova m1, m5
2002 movu m1, [lq+mmsize*0+2]
2005 pavgw m1, m2
2025 SHIFT_RIGHTx2 m1, m6, m7, m0
2027 pavgw m7, m1
2029 pshufd m1, m1, q3333
2043 mova [rsp+1*mmsize], m1
2044 mova m1, [rsp+0*mmsize]
2046 mova [dstq+strideq *0+ 0], m1
2059 vpalignr m1, m0, m1, 4
2067 PALIGNR m6, m0, m1, 4, m7
2068 mova m1, m6
2083 mova [rsp+0*mmsize], m1
2084 mova m1, [rsp+1*mmsize]
2097 mova [dstq+stride20q +48], m1
2100 mova [dstq+stride3q*8+32], m1
2101 mova [dstq+stride3q*8+48], m1
2103 mova [dstq+stride28q +16], m1
2104 mova [dstq+stride28q +32], m1
2105 mova [dstq+stride28q +48], m1
2111 vpalignr m6, m1, m6, 4
2121 PALIGNR m0, m1, m6, 4, m3
2145 psrldq m1, m0, 2
2148 pavgw m1, m0
2149 punpcklwd m1, m2
2153 movh [dstq+stride3q ], m1
2154 movhps [dstq+strideq*1], m1
2156 PALIGNR m2, m1, 4, m0
2163 movu m1, [aq-2]
2164 PALIGNR m2, m1, m0, 2, m3
2165 PALIGNR m3, m1, m0, 4, m4
2169 psrldq m0, m1, 2
2170 psrldq m4, m1, 4
2183 vpalignr m3, m1, m3, 4
2187 PALIGNR m0, m1, m3, 4, m4
2190 psrldq m1, 4
2197 movu m1, [lq+2]
2200 pavgw m1, m2
2231 mova [dstq+mstrideq *0+ 0], m1
2234 vpalignr m1, m0, m1, 4
2240 PALIGNR m6, m0, m1, 4, m7
2241 mova m1, m6
2259 movu m1, [lq+mmsize*0+2]
2262 pavgw m1, m2
2281 movu m1, [aq+mmsize*0-2]
2282 PALIGNR m7, m1, m0, 2, m2
2283 PALIGNR m6, m1, m0, 4, m2
2290 movu m1, [aq+mmsize*1-2]
2355 mova [dstq+strideq* 4+48], m1
2358 mova [dstq+strideq* 0+32], m1
2372 vpalignr m0, m1, m0, 4
2373 vpalignr m1, m2, m1, 4
2398 PALIGNR m3, m1, m0, 4, m2
2404 PALIGNR m3, m2, m1, 4, m0
2405 mova m1, m3
2432 mova [dstq+stride3q*4+ 0], m1
2450 vpalignr m1, m0, m1, 4
2462 PALIGNR m6, m0, m1, 4, m7
2463 mova m1, m6