• Home
  • Raw
  • Download

Lines Matching refs:T1

420 my ($H0,$H1,$H2,$H3,$H4, $T0,$T1,$T2,$T3,$T4, $D0,$D1,$D2,$D3,$D4, $MASK) =
888 vmovdqu 16*3($inp),$T1
892 vpsrldq \$6,$T1,$T3
893 vpunpckhqdq $T1,$T0,$T4 # 4
894 vpunpcklqdq $T1,$T0,$T0 # 0:1
898 vpsrlq \$26,$T0,$T1
901 vpand $MASK,$T1,$T1 # 1
981 vpmuludq $T1,$D4,$D1 # d1 = h1*r0
996 vpmuludq $T1,$H2,$H1 # h1*r1
1007 vpmuludq $T1,$H3,$H1 # h1*r2
1019 vpmuludq $T1,$H2,$H1 # h1*r3
1041 vpmuludq $T1,$H4,$T0 # h1*s4
1083 vpmuludq $H1,$T4,$T1 # h1*r0
1085 vpaddq $T1,$D1,$D1
1088 vpmuludq $H3,$T4,$T1 # h3*r0
1090 vpaddq $T1,$D3,$D3
1096 vpmuludq $H2,$T2,$T1 # h2*r1
1098 vpaddq $T1,$D3,$D3 # d3 += h2*r1
1101 vpmuludq $H1,$T2,$T1 # h1*r1
1103 vpaddq $T1,$D2,$D2 # d2 += h1*r1
1108 vpmuludq $H1,$T3,$T1 # h1*r2
1110 vpaddq $T1,$D3,$D3 # d3 += h1*r2
1118 vpmuludq $H1,$T2,$T1 # h1*r3
1122 vpaddq $T1,$D4,$D4 # d4 += h1*r3
1131 vmovdqu 16*3($inp),$T1 #
1139 vpsrldq \$6,$T1,$T3 #
1144 vpunpckhqdq $T1,$T0,$T4 # 4
1148 vpunpcklqdq $T1,$T0,$T0 # 0:1
1153 vpsrlq \$26,$T0,$T1
1157 vpand $MASK,$T1,$T1 # 1
1211 vpaddq $H1,$T1,$T1
1231 vpmuludq $T1,$D4,$D1 # d1 = h1*r0
1241 vpmuludq $T1,$H2,$H0 # h1*r1
1251 vpmuludq $T1,$H4,$H0 # h1*r2
1262 vpmuludq $T1,$H3,$H0 # h1*r3
1283 vpmuludq $T1,$H3,$H3 # h1*s4
1319 vpmuludq $H1,$T4,$T1 # h1*r0
1320 vpaddq $T1,$D1,$D1 # d1 += h1*r0
1324 vpmuludq $H3,$T4,$T1 # h3*r0
1325 vpaddq $T1,$D3,$D3 # d3 += h3*r0
1332 vpmuludq $H2,$T2,$T1 # h2*r1
1333 vpaddq $T1,$D3,$D3 # d3 += h2*r1
1343 vpmuludq $H2,$T4,$T1 # h2*r2
1344 vpaddq $T1,$D4,$D4 # d4 += h2*r2
1350 vpmuludq $H4,$T2,$T1 # h4*s2
1351 vpaddq $T1,$D1,$D1 # d1 += h4*s2
1361 vpmuludq $H4,$T4,$T1 # h4*s3
1362 vpaddq $T1,$D2,$D2 # d2 += h4*s3
1371 vpmuludq $H4,$T3,$T1 # h4*s4
1372 vpaddq $T1,$D3,$D3 # d3 += h4*s4
1375 vpmuludq $H2,$T3,$T1 # h2*s4
1376 vpaddq $T1,$D1,$D1 # d1 += h2*s4
1386 vpsrldq \$8,$D1,$T1
1392 vpaddq $T1,$D1,$D1
1517 my ($H0,$H1,$H2,$H3,$H4, $MASK, $T4,$T0,$T1,$T2,$T3, $D0,$D1,$D2,$D3,$D4) =
1869 vmovdqu 16*1($inp),%x#$T1
1871 vinserti128 \$1,16*3($inp),$T1,$T1
1875 vpsrldq \$6,$T1,$T3
1876 vpunpckhqdq $T1,$T0,$T4 # 4
1878 vpunpcklqdq $T1,$T0,$T0 # 0:1
1882 vpsrlq \$26,$T0,$T1
1886 vpand $MASK,$T1,$T1 # 1
1907 vpaddq $H1,$T1,$H1
1908 vmovdqa `32*1`(%rsp),$T1 # r1^4
1931 vpmuludq $H2,$T1,$D3 # d3 = h2*r1
1936 vpmuludq $H0,$T1,$T4 # h0*r1
1937 vpmuludq $H1,$T1,$H2 # h1*r1, borrow $H2 as temp
1940 vpmuludq $H3,$T1,$T4 # h3*r1
1944 vmovdqa `32*4-0x90`(%rax),$T1 # s2
1957 vpmuludq $H3,$T1,$T4 # h3*s2
1958 vpmuludq $H4,$T1,$H2 # h4*s2
1959 vmovdqu 16*1($inp),%x#$T1
1967 vinserti128 \$1,16*3($inp),$T1,$T1
1977 vpsrldq \$6,$T1,$T3
1980 vpunpckhqdq $T1,$T0,$T4 # 4
1984 vpunpcklqdq $T1,$T0,$T0 # 0:1
2019 vpsrlq \$26,$T0,$T1
2039 vpand $MASK,$T1,$T1 # 1
2057 vpaddq $H1,$T1,$H1
2058 vmovdqu `32*1+4`(%rsp),$T1 # r1^4
2066 vpmuludq $H2,$T1,$D3 # d3 = h2*r1
2071 vpmuludq $H0,$T1,$T4 # h0*r1
2072 vpmuludq $H1,$T1,$H2 # h1*r1
2075 vpmuludq $H3,$T1,$T4 # h3*r1
2083 vmovdqu `32*4+4-0x90`(%rax),$T1 # s2
2090 vpmuludq $H3,$T1,$T4 # h3*s2
2091 vpmuludq $H4,$T1,$H2 # h4*s2
2122 vpsrldq \$8,$D1,$T1
2127 vpaddq $T1,$D1,$D1
2136 vpermq \$0x2,$D1,$T1
2141 vpaddq $T1,$D1,$D1
2212 map(s/%y/%z/,($T4,$T0,$T1,$T2,$T3)); # switch to %zmm domain
2255 vmovdqu `16*4-64`($ctx),%x#$T1 # ... ${S2}
2267 vpermd $T1,$T2,$S2
2269 vpsrlq \$32,$R1,$T1
2298 vpmuludq $T1,$S4,$M0
2299 vpmuludq $T1,$R0,$M1
2300 vpmuludq $T1,$R1,$M2
2301 vpmuludq $T1,$R2,$M3
2302 vpmuludq $T1,$R3,$M4
2426 vpsrlq \$26,$T0,$T1
2431 #vpandq $MASK,$T1,$T1 # 1
2473 vpandq $MASK,$T1,$T1 # 1
2479 vpaddq $H1,$T1,$H1 # accumulate input
2564 vpsrlq \$26,$T0,$T1
2583 #vpandq $MASK,$T1,$T1 # 1
2616 vpandq $MASK,$T1,$T1 # 1
2621 vpaddq $H1,$T1,$H1 # accumulate input
2635 vmovdqu 16*1($inp),%x#$T1
2655 vinserti128 \$1,16*3($inp),%y#$T1,%y#$T1
2712 map(s/%z/%y/,($T0,$T1,$T2,$T3,$T4, $PADBIT));
2721 vpsrldq \$6,$T1,$T3
2722 vpunpckhqdq $T1,$T0,$T4 # 4
2728 vpunpcklqdq $T1,$T0,$T0 # 0:1
2742 vpsrlq \$26,$T0,$T1
2755 vpand $MASK,$T1,$T1 # 1
3025 my ($T0,$T1,$T2,$T3,$mask44,$mask42,$tmp,$PAD) = map("%ymm$_",(24..31));
3073 vpunpcklqdq $T3,$T2,$T1 # transpose data
3081 vpandq $mask44,$T1,$T0
3082 vpsrlq \$44,$T1,$T1
3084 vporq $T3,$T1,$T1
3085 vpandq $mask44,$T1,$T1
3243 vpunpcklqdq $T3,$T2,$T1 # transpose data
3251 vpandq $mask44,$T1,$T0
3252 vpsrlq \$44,$T1,$T1
3254 vporq $T3,$T1,$T1
3255 vpandq $mask44,$T1,$T1
3264 vpaddq $T1,$H1,$H1
3289 vpunpcklqdq $T3,$T2,$T1 # transpose data
3314 vpandq $mask44,$T1,$T0
3315 vpsrlq \$44,$T1,$T1
3329 vporq $T3,$T1,$T1
3330 vpandq $mask44,$T1,$T1
3353 vpaddq $T1,$H1,$H1
3389 vpsrldq \$8,$D1lo,$T1
3395 vpaddq $T1,$D1lo,$D1lo
3402 vpermq \$0x2,$D1lo,$T1
3408 vpaddq $T1,$D1lo,${D1lo}{%k1}{z}
3466 my ($T0,$T1,$T2,$T3,$mask44,$mask42,$tmp,$PAD) = map("%ymm$_",(24..31));
3572 vpunpcklqdq $R1,$RR1,$T1
3578 map(s/%y/%z/, $T0,$T1,$T2,$T3,$mask44,$mask42,$tmp,$PAD);
3584 vshufi64x2 \$0x44,$R1,$T1,$RR1
3607 vpunpcklqdq $T3,$T2,$T1 # transpose data
3615 vpandq $mask44,$T1,$T0
3616 vpsrlq \$44,$T1,$T1
3618 vporq $T3,$T1,$T1
3619 vpandq $mask44,$T1,$T1
3629 vpaddq $T1,$H1,$H1
3654 vpunpcklqdq $T3,$T2,$T1 # transpose data
3679 vpandq $mask44,$T1,$T0
3680 vpsrlq \$44,$T1,$T1
3694 vporq $T3,$T1,$T1
3695 vpandq $mask44,$T1,$T1
3708 vpaddq $T1,$H1,$H1
3744 vpsrldq \$8,$D1lo,$T1
3750 vpaddq $T1,$D1lo,$D1lo
3757 vpermq \$0x2,$D1lo,$T1
3763 vpaddq $T1,$D1lo,$D1lo
3770 vextracti64x4 \$1,$D1lo,%y#$T1
3778 map(s/%z/%y/, $T0,$T1,$T2,$T3,$mask44,$mask42,$tmp,$PAD);
3783 vpaddq $T1,$D1lo,${D1lo}{%k1}{z}