Lines Matching refs:H1
419 my ($H0,$H1,$H2,$H3,$H4, $T0,$T1,$T2,$T3,$T4, $D0,$D1,$D2,$D3,$D4, $MASK) =
713 vmovd %rdx#d,$H1
815 vmovd %rdx#d,$H1
845 vmovd 4*1($ctx),$H1
989 vmovdqa $H1,0x10(%r11) #
990 vpmuludq $T3,$H2,$H1 # h3*r1
992 vpaddq $H1,$D4,$D4 # d4 += h3*r1
995 vpmuludq $T1,$H2,$H1 # h1*r1
998 vpaddq $H1,$D2,$D2 # d2 += h1*r1
1006 vpmuludq $T1,$H3,$H1 # h1*r2
1008 vpaddq $H1,$D3,$D3 # d3 += h1*r2
1018 vpmuludq $T1,$H2,$H1 # h1*r3
1020 vpaddq $H1,$D4,$D4 # d4 += h1*r3
1023 vpmuludq $T3,$H3,$H1 # h3*s3
1026 vpaddq $H1,$D1,$D1 # d1 += h3*s3
1031 vmovdqu 16*1($inp),$H1 #
1038 vpsrldq \$6,$H1,$H3 #
1041 vpunpckhqdq $H1,$H0,$H4 # 4
1046 vpunpcklqdq $H1,$H0,$H0 # 0:1
1051 vpsrlq \$26,$H0,$H1
1054 vpand $MASK,$H1,$H1 # 1
1062 vpaddq 0x10(%r11),$H1,$H1
1082 vpmuludq $H1,$T4,$T1 # h1*r0
1100 vpmuludq $H1,$T2,$T1 # h1*r1
1107 vpmuludq $H1,$T3,$T1 # h1*r2
1117 vpmuludq $H1,$T2,$T1 # h1*r3
1142 vpmuludq $H1,$T4,$H0
1173 vpaddq $D0,$D1,$H1 # h0 -> h1
1178 vpsrlq \$26,$H1,$D1
1179 vpand $MASK,$H1,$H1
1192 vpaddq $D0,$H1,$H1 # h0 -> h1
1210 vpaddq $H1,$T1,$T1
1217 vmovdqa $H1,0x10(%r11)
1237 vpmuludq $T2,$H2,$H1 # h2*r1
1238 vpaddq $H1,$D3,$D3 # d3 += h2*r1
1248 vpmuludq $T2,$H4,$H1 # h2*r2
1249 vpaddq $H1,$D4,$D4 # d4 += h2*r2
1255 vpmuludq $T4,$H2,$H1 # h4*s2
1256 vpaddq $H1,$D1,$D1 # d1 += h4*s2
1266 vpmuludq $T4,$H4,$H1 # h4*s3
1267 vpaddq $H1,$D2,$D2 # d2 += h4*s3
1276 vpmuludq $T4,$H3,$H1 # h4*s4
1277 vpaddq $H1,$D3,$D3 # h3 = d3 + h4*s4
1280 vpmuludq $T2,$H3,$H1 # h2*s4
1281 vpaddq $H1,$D1,$D1 # h1 = d1 + h2*s4
1288 vmovdqu 16*1($inp),$H1
1291 vpsrldq \$6,$H1,$H3
1292 vpunpckhqdq $H1,$H0,$H4 # 4
1293 vpunpcklqdq $H1,$H0,$H0 # 0:1
1297 vpsrlq \$26,$H0,$H1
1300 vpand $MASK,$H1,$H1 # 1
1308 vpaddq 0x10(%r11),$H1,$H1
1318 vpmuludq $H1,$T4,$T1 # h1*r0
1334 vpmuludq $H1,$T2,$T0 # h1*r1
1344 vpmuludq $H1,$T4,$T0 # h1*r2
1355 vpmuludq $H1,$T3,$T0 # h1*r3
1376 vpmuludq $H1,$T3,$T3 # h1*s4
1408 vpsrlq \$26,$D1,$H1
1410 vpaddq $H1,$D2,$D2 # h1 -> h2
1516 my ($H0,$H1,$H2,$H3,$H4, $MASK, $T4,$T0,$T1,$T2,$T3, $D0,$D1,$D2,$D3,$D4) =
1641 vmovd %rdx#d,%x#$H1
1748 vmovd %rdx#d,%x#$H1
1789 vmovd 4*1($ctx),%x#$H1
1906 vpaddq $H1,$T1,$H1
1936 vpmuludq $H1,$T1,$H2 # h1*r1, borrow $H2 as temp
1946 vpmuludq $H1,$T0,$H2 # h1*r0
1962 vpmuludq $H1,$T2,$T4 # h1*r2
1969 vpmuludq $H1,$H2,$T4 # h1*r3
1988 vpmuludq $H1,$S4,$H0 # h1*s4
2002 vpaddq $D0,$D1,$H1 # h0 -> h1
2009 vpsrlq \$26,$H1,$D1
2010 vpand $MASK,$H1,$H1
2029 vpaddq $D0,$H1,$H1 # h0 -> h1
2056 vpaddq $H1,$T1,$H1
2071 vpmuludq $H1,$T1,$H2 # h1*r1
2080 vpmuludq $H1,$T0,$H2 # h1*r0
2094 vpmuludq $H1,$T2,$T4 # h1*r2
2099 vpmuludq $H1,$H2,$T4 # h1*r3
2113 vpmuludq $H1,$S4,$H0 # h1*s4
2152 vpaddq $D0,$D1,$H1 # h0 -> h1
2157 vpsrlq \$26,$H1,$D1
2158 vpand $MASK,$H1,$H1
2171 vpaddq $D0,$H1,$H1 # h0 -> h1
2178 vmovd %x#$H1,`4*1-48-64`($ctx)
2213 map(s/%y/%z/,($H0,$H1,$H2,$H3,$H4));
2478 vpaddq $H1,$T1,$H1 # accumulate input
2494 vpmuludq $H1,$R2,$M3
2495 vpmuludq $H1,$R3,$M4
2496 vpmuludq $H1,$S4,$M0
2508 vpmuludq $H1,$R0,$M1
2509 vpmuludq $H1,$R1,$M2
2530 vpaddq $M1,$D1,$H1 # h1 = d2 + h4*s2
2547 vpaddq $D0,$H1,$H1 # h0 -> h1
2554 vpsrlq \$26,$H1,$D1
2555 vpandq $MASK,$H1,$H1
2573 vpaddq $D0,$H1,$H1 # h0 -> h1
2620 vpaddq $H1,$T1,$H1 # accumulate input
2635 vpmuludq $H1,$R2,$M3
2636 vpmuludq $H1,$R3,$M4
2637 vpmuludq $H1,$S4,$M0
2647 vpmuludq $H1,$R0,$M1
2648 vpmuludq $H1,$R1,$M2
2670 vpaddq $M1,$D1,$H1 # h1 = d2 + h4*s2
2680 vpermq \$0xb1,$H1,$D1
2685 vpaddq $D1,$H1,$H1
2692 vpermq \$0x2,$H1,$D1
2697 vpaddq $D1,$H1,$H1
2703 vextracti64x4 \$0x1,$H1,%y#$D1
2708 vpaddq $D1,$H1,${H1}{%k3}{z}
2712 map(s/%z/%y/,($H0,$H1,$H2,$H3,$H4, $D0,$D1,$D2,$D3,$D4, $MASK));
2728 vpaddq $D0,$H1,$H1 # h0 -> h1
2733 vpsrlq \$26,$H1,$D1
2734 vpand $MASK,$H1,$H1
2755 vpaddq $D0,$H1,$H1 # h0 -> h1
2769 vmovd %x#$H1,`4*1-48-64`($ctx)
2899 my ($H0,$H1,$H2,$r2r1r0,$r1r0s2,$r0s2s1,$Dlo,$Dhi) = map("%ymm$_",(0..5,16,17));
2964 vpermq \$0b01010101,$Dlo,${H1}{%k7}{z}
2973 vpmadd52luq $r1r0s2,$H1,$Dlo
2974 vpmadd52huq $r1r0s2,$H1,$Dhi
3022 my ($H0,$H1,$H2,$R0,$R1,$R2,$S1,$S2) = map("%ymm$_",(0..5,16,17));
3048 vmovq 8($ctx),%x#$H1
3100 vmovdqa $R1,$H1
3126 vpmadd52luq $H1,$S2,$D0lo
3127 vpmadd52huq $H1,$S2,$D0hi
3128 vpmadd52luq $H1,$R0,$D1lo
3129 vpmadd52huq $H1,$R0,$D1hi
3130 vpmadd52luq $H1,$R1,$D2lo
3131 vpmadd52huq $H1,$R1,$D2hi
3144 vpandq $mask44,$D1lo,$H1
3162 vpaddq $tmp,$H1,$H1
3167 vpunpcklqdq $R1,$H1,$R1 # 1,2
3168 vpbroadcastq %x#$H1,%x#$H1 # 2,2
3186 vinserti128 \$1,%x#$R1,$H1,$R1 # 1,2,3,4
3199 vmovq 8($ctx),%x#$H1
3263 vpaddq $T1,$H1,$H1
3290 vpmadd52luq $H1,$S2,$D0lo
3291 vpmadd52huq $H1,$S2,$D0hi
3292 vpmadd52luq $H1,$R0,$D1lo
3293 vpmadd52huq $H1,$R0,$D1hi
3294 vpmadd52luq $H1,$R1,$D2lo
3295 vpmadd52huq $H1,$R1,$D2hi
3310 vpandq $mask44,$D1lo,$H1
3334 vpaddq $tmp,$H1,$H1
3352 vpaddq $T1,$H1,$H1
3374 vpmadd52luq $H1,$S2,$D0lo
3375 vpmadd52huq $H1,$S2,$D0hi
3376 vpmadd52luq $H1,$R0,$D1lo
3377 vpmadd52huq $H1,$R0,$D1hi
3378 vpmadd52luq $H1,$R1,$D2lo
3379 vpmadd52huq $H1,$R1,$D2hi
3389 vpsrldq \$8,$D1hi,$H1
3395 vpaddq $H1,$D1hi,$D1hi
3402 vpermq \$0x2,$D1hi,$H1
3408 vpaddq $H1,$D1hi,${D1hi}{%k1}{z}
3423 vpandq $mask44,$D1lo,$H1
3441 vpaddq $tmp,$H1,$H1
3448 vmovq %x#$H1,8($ctx)
3463 my ($H0,$H1,$H2,$R0,$R1,$R2,$S1,$S2) = map("%ymm$_",(0..5,16,17));
3485 vmovq 8($ctx),%x#$H1
3575 map(s/%y/%z/, $H0,$H1,$H2,$R0,$R1,$R2,$S1,$S2);
3628 vpaddq $T1,$H1,$H1
3655 vpmadd52luq $H1,$S2,$D0lo
3656 vpmadd52huq $H1,$S2,$D0hi
3657 vpmadd52luq $H1,$R0,$D1lo
3658 vpmadd52huq $H1,$R0,$D1hi
3659 vpmadd52luq $H1,$R1,$D2lo
3660 vpmadd52huq $H1,$R1,$D2hi
3675 vpandq $mask44,$D1lo,$H1
3699 vpaddq $tmp,$H1,$H1
3707 vpaddq $T1,$H1,$H1
3729 vpmadd52luq $H1,$SS2,$D0lo
3730 vpmadd52huq $H1,$SS2,$D0hi
3731 vpmadd52luq $H1,$RR0,$D1lo
3732 vpmadd52huq $H1,$RR0,$D1hi
3733 vpmadd52luq $H1,$RR1,$D2lo
3734 vpmadd52huq $H1,$RR1,$D2hi
3744 vpsrldq \$8,$D1hi,$H1
3750 vpaddq $H1,$D1hi,$D1hi
3757 vpermq \$0x2,$D1hi,$H1
3763 vpaddq $H1,$D1hi,$D1hi
3770 vextracti64x4 \$1,$D1hi,%y#$H1
3775 map(s/%z/%y/, $H0,$H1,$H2,$R0,$R1,$R2,$S1,$S2);
3783 vpaddq $H1,$D1hi,${D1hi}{%k1}{z}
3798 vpandq $mask44,$D1lo,$H1
3816 vpaddq $tmp,$H1,$H1
3821 vmovq %x#$H1,8($ctx)