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
1228 mova m1, [lq+mmsize*1+0] ; l[16-31]
1232 vperm2i128 m5, m0, m1, q0201 ; lmnopqrstuvwxyz0
1236 vperm2i128 m7, m1, m2, q0201 ; stuvwxyz*abcdefg
1237 vpalignr m5, m7, m1, 2 ; lmnopqrstuvwxyz*
1238 vpalignr m6, m7, m1, 4 ; mnopqrstuvwxyz*a
1247 vperm2i128 m4, m1, m2, q0201 ; TUVWXYZ#ABCDEFGH
1248 vperm2i128 m5, m0, m1, q0201 ; L[7-15]L[16-23]
1260 mova [dst24q+stride7q+32], m1
1261 mova [dst8q+stride7q+0], m1
1263 vpalignr m6, m4, m1, 2
1270 vpalignr m6, m4, m1, 4
1277 vpalignr m6, m4, m1, 6
1284 vpalignr m6, m4, m1, 8
1291 vpalignr m6, m4, m1, 10
1298 vpalignr m6, m4, m1, 12
1305 vpalignr m6, m4, m1, 14
1313 mova m5, m1
1314 mova m1, m4
1332 psrldq m1, m0, 2 ; bcdefgh.
1335 pavgw m1, m0 ; ABCDEFG.
1339 movh [dstq+strideq*0], m1
1341 psrldq m1, 2
1343 movh [dstq+strideq*2], m1
1353 SHIFT_RIGHTx2 m1, m2, m0, m3 ; bcdefghh/cdefghhh
1355 pavgw m1, m0 ; ABCDEFGh
1359 mova [dstq+strideq*0], m1
1361 SHIFT_RIGHT m1, m1, m3
1363 mova [dstq+strideq*2], m1
1366 SHIFT_RIGHT m1, m1, m3
1368 mova [dstq+strideq*0], m1
1370 SHIFT_RIGHT m1, m1, m3
1372 mova [dstq+strideq*2], m1
1379 mova m1, [aq+mmsize]
1380 PALIGNR m2, m1, m0, 2, m3
1381 PALIGNR m3, m1, m0, 4, m4
1387 SHIFT_RIGHTx2 m5, m0, m1, m4
1389 pavgw m1, m5
1395 mova [dstq+strideq*0+16], m1
1400 vpalignr m2, m1, m2, 2
1403 PALIGNR m5, m1, m2, 2, m4
1408 SHIFT_RIGHT m1, m1, m4
1417 mova m1, [aq+mmsize*1]
1419 PALIGNR m6, m1, m0, 2, m5
1420 PALIGNR m7, m1, m0, 4, m5
1424 PALIGNR m4, m2, m1, 2, m0
1425 PALIGNR m5, m2, m1, 4, m0
1427 pavgw m4, m1
1429 PALIGNR m1, m0, m2, 2, m6
1432 pavgw m2, m1
1436 SHIFT_RIGHTx2 m6, m1, m0, reg_shuf
1463 mova [dstq+strideq*1+48], m1
1472 mova [dstq+stride17q+32], m1
1483 vpalignr m3, m1, m3, 2
1489 PALIGNR m3, m4, m6, 2, m1
1491 PALIGNR m3, m2, m4, 2, m1
1493 PALIGNR m3, m0, m2, 2, m1
1495 PALIGNR m3, m5, m7, 2, m1
1505 PALIGNR m6, m1, m3, 2, m7
1511 SHIFT_RIGHT m1, m1, reg_shuf
1544 movhps m1, [lq]
1545 PALIGNR m0, m1, 10, m2 ; xyz*abcd
1546 pslldq m1, m0, 2 ; .xyz*abc
1549 pavgw m1, m0 ; ....#ABC
1553 movhps [dstq+strideq*0], m1
1555 shufps m0, m2, m1, q3210
1568 movu m1, [aq-2] ; *abcdefg
1571 PALIGNR m3, m1, m2, 14, m4 ; z*abcdef
1573 pavgw m0, m1
1574 PALIGNR m1, m2, 2, m4 ; tuvwxyz*
1582 PALIGNR m0, m4, 14, m1
1584 PALIGNR m3, m4, 14, m1
1589 PALIGNR m0, m4, 14, m1
1591 PALIGNR m3, m4, 14, m1
1595 PALIGNR m0, m4, 14, m1
1603 movu m1, [aq-2] ; *abcdefg
1608 PALIGNR m0, m1, m5, 14, m6 ; z*abcdef
1613 pavgw m3, m1
1614 PALIGNR m1, m5, 2, m7 ; tuvwxyz*
1621 psrld m4, m1, 16
1623 pand m1, [pd_65535]
1626 packssdw m5, m1
1648 movu m1, [aq+mmsize*1-2] ; a[7-14]
1665 pavgw m1, m4
1712 mova [dstq+strideq*0+16], m1
1723 mova [dstq+stride16q+32], m1
1733 PALIGNR m2, m1, 14, m5
1734 PALIGNR m1, m4, 14, m5
1762 mova [dstq+strideq*0+48], m1
1768 PALIGNR m1, m2, 14, m6
1796 psrldq m1, m0, 2 ; bcddddd.
1799 pavgw m1, m0 ; abcddddd
1801 PALIGNR m2, m1, 4, m0 ; bCcDdddd
1805 movh [dstq+strideq*0], m1 ; aBbC
1807 movhps [dstq+strideq*2], m1 ; cDdd
1816 SHIFT_RIGHTx2 m1, m2, m0, m3
1818 pavgw m1, m0
1820 shufps m0, m1, m2, q1032
1825 mova [dstq+strideq *0], m1
1831 vpalignr m1, m2, m1, 4
1833 PALIGNR m0, m2, m1, 4, m3
1834 mova m1, m0
1837 shufps m0, m1, m2, q1032
1839 mova [dstq+strideq *0], m1
1848 movu m1, [lq+2]
1851 pavgw m1, m0
1866 mova [dstq+strideq *0+ 0], m1
1876 vpalignr m1, m2, m1, 4
1881 PALIGNR m5, m2, m1, 4, m6
1882 mova m1, m5
1897 movu m1, [lq+mmsize*0+2]
1900 pavgw m1, m2
1920 SHIFT_RIGHTx2 m1, m6, m7, m0
1922 pavgw m7, m1
1924 pshufd m1, m1, q3333
1938 mova [rsp+1*mmsize], m1
1939 mova m1, [rsp+0*mmsize]
1941 mova [dstq+strideq *0+ 0], m1
1954 vpalignr m1, m0, m1, 4
1962 PALIGNR m6, m0, m1, 4, m7
1963 mova m1, m6
1978 mova [rsp+0*mmsize], m1
1979 mova m1, [rsp+1*mmsize]
1992 mova [dstq+stride20q +48], m1
1995 mova [dstq+stride3q*8+32], m1
1996 mova [dstq+stride3q*8+48], m1
1998 mova [dstq+stride28q +16], m1
1999 mova [dstq+stride28q +32], m1
2000 mova [dstq+stride28q +48], m1
2006 vpalignr m6, m1, m6, 4
2016 PALIGNR m0, m1, m6, 4, m3
2040 psrldq m1, m0, 2
2043 pavgw m1, m0
2044 punpcklwd m1, m2
2048 movh [dstq+stride3q ], m1
2049 movhps [dstq+strideq*1], m1
2051 PALIGNR m2, m1, 4, m0
2058 movu m1, [aq-2]
2059 PALIGNR m2, m1, m0, 2, m3
2060 PALIGNR m3, m1, m0, 4, m4
2064 psrldq m0, m1, 2
2065 psrldq m4, m1, 4
2078 vpalignr m3, m1, m3, 4
2082 PALIGNR m0, m1, m3, 4, m4
2085 psrldq m1, 4
2092 movu m1, [lq+2]
2095 pavgw m1, m2
2126 mova [dstq+mstrideq *0+ 0], m1
2129 vpalignr m1, m0, m1, 4
2135 PALIGNR m6, m0, m1, 4, m7
2136 mova m1, m6
2154 movu m1, [lq+mmsize*0+2]
2157 pavgw m1, m2
2176 movu m1, [aq+mmsize*0-2]
2177 PALIGNR m7, m1, m0, 2, m2
2178 PALIGNR m6, m1, m0, 4, m2
2185 movu m1, [aq+mmsize*1-2]
2250 mova [dstq+strideq* 4+48], m1
2253 mova [dstq+strideq* 0+32], m1
2267 vpalignr m0, m1, m0, 4
2268 vpalignr m1, m2, m1, 4
2293 PALIGNR m3, m1, m0, 4, m2
2299 PALIGNR m3, m2, m1, 4, m0
2300 mova m1, m3
2327 mova [dstq+stride3q*4+ 0], m1
2345 vpalignr m1, m0, m1, 4
2357 PALIGNR m6, m0, m1, 4, m7
2358 mova m1, m6