#
# Optimization including one of Pavel Semjanov's ideas, alternative
# Maj, resulted in >=5% improvement on most CPUs, +20% SHA256 and
-# unfortunately -10% SHA512 on P4 [which nobody should care about
+# unfortunately -2% SHA512 on P4 [which nobody should care about
# that much].
#
# June 2012.
#
# SHA256 SSSE3 AVX/XOP(*) SHA512 AVX/XOP(*)
#
-# AMD K8 15.1 - - 9.70 -
-# P4 17.5 - - 33.4 -
-# Core 2 15.5 13.9(+11%) - 10.3 -
-# Westmere 15.1 12.5(+21%) - 9.72 -
-# Sandy Bridge 17.4 14.0(+24%) 11.6(+50%(**)) 11.2 8.10(+38%(**))
-# Ivy Bridge 12.6 10.3(+22%) 10.3(+22%) 8.17 7.22(+13%)
-# Bulldozer 21.5 13.7(+57%) 13.7(+57%(***)) 13.5 8.58(+57%)
-# VIA Nano 23.0 16.3(+41%) - 14.7 -
-# Atom 23.0 21.6(+6%) - 14.7 -
+# AMD K8 14.9 - - 9.57 -
+# P4 17.3 - - 30.8 -
+# Core 2 15.6 13.8(+13%) - 9.97 -
+# Westmere 14.8 12.3(+19%) - 9.58 -
+# Sandy Bridge 17.4 14.2(+23%) 11.6(+50%(**)) 11.2 8.10(+38%(**))
+# Ivy Bridge 12.6 10.5(+20%) 10.3(+22%) 8.17 7.22(+13%)
+# Haswell 12.2 9.28(+31%) 7.80(+56%) 7.66 5.40(+42%)
+# Bulldozer 21.1 13.6(+54%) 13.6(+54%(***)) 13.5 8.58(+57%)
+# VIA Nano 23.0 16.5(+39%) - 14.7 -
+# Atom 23.0 18.9(+22%) - 14.7 -
#
# (*) whichever best applicable;
# (**) switch from ror to shrd stands for fair share of improvement;
ror \$`$Sigma1[2]-$Sigma1[1]`,$a0
mov $f,$a2
- ror \$`$Sigma0[2]-$Sigma0[1]`,$a1
xor $e,$a0
+ ror \$`$Sigma0[2]-$Sigma0[1]`,$a1
xor $g,$a2 # f^g
mov $T1,`$SZ*($i&0xf)`(%rsp)
add ($Tbl),$T1 # T1+=K[round]
xor $a,$a1
- ror \$$Sigma1[0],$a0 # Sigma1(e)
xor $b,$a2 # a^b, b^c in next round
+ ror \$$Sigma1[0],$a0 # Sigma1(e)
mov $b,$h
- ror \$$Sigma0[0],$a1 # Sigma0(a)
and $a2,$a3
+ ror \$$Sigma0[0],$a1 # Sigma0(a)
add $a0,$T1 # T1+=Sigma1(e)
xor $a3,$h # h=Maj(a,b,c)=Ch(a^b,c,b)
add $T1,$d # d+=T1
add $T1,$h # h+=T1
-___
-$code.=<<___ if ($i>=15);
- mov `$SZ*(($i+2)&0xf)`(%rsp),$a0
-___
-$code.=<<___;
+
lea $STRIDE($Tbl),$Tbl # round++
+___
+$code.=<<___ if ($i<15);
add $a1,$h # h+=Sigma0(a)
-
___
($a2,$a3) = ($a3,$a2);
}
{ my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
$code.=<<___;
- #mov `$SZ*(($i+1)&0xf)`(%rsp),$a0
- mov `$SZ*(($i+14)&0xf)`(%rsp),$a1
+ mov `$SZ*(($i+1)&0xf)`(%rsp),$a0
+ mov `$SZ*(($i+14)&0xf)`(%rsp),$a2
mov $a0,$T1
ror \$`$sigma0[1]-$sigma0[0]`,$a0
- mov $a1,$a2
- ror \$`$sigma1[1]-$sigma1[0]`,$a1
+ add $a1,$a # modulo-scheduled h+=Sigma0(a)
+ mov $a2,$a1
+ ror \$`$sigma1[1]-$sigma1[0]`,$a2
xor $T1,$a0
shr \$$sigma0[2],$T1
ror \$$sigma0[0],$a0
- xor $a2,$a1
- shr \$$sigma1[2],$a2
+ xor $a1,$a2
+ shr \$$sigma1[2],$a1
+ ror \$$sigma1[0],$a2
xor $a0,$T1 # sigma0(X[(i+1)&0xf])
- ror \$$sigma1[0],$a1
+ xor $a1,$a2 # sigma1(X[(i+14)&0xf])
add `$SZ*(($i+9)&0xf)`(%rsp),$T1
- xor $a2,$a1 # sigma1(X[(i+14)&0xf])
add `$SZ*($i&0xf)`(%rsp),$T1
mov $e,$a0
- add $a1,$T1
+ add $a2,$T1
mov $a,$a1
___
&ROUND_00_15(@_);
jnz .Lrounds_16_xx
mov $_ctx,$ctx
+ add $a1,$A # modulo-scheduled h+=Sigma0(a)
lea 16*$SZ($inp),$inp
add $SZ*0($ctx),$A
'&mov ($a,$a1)',
'&mov ($a4,$f)',
- '&xor ($a0,$e)',
'&ror ($a1,$Sigma0[2]-$Sigma0[1])',
+ '&xor ($a0,$e)',
'&xor ($a4,$g)', # f^g
'&ror ($a0,$Sigma1[1]-$Sigma1[0])',
'&add ($h,$SZ*($i&15)."(%rsp)")', # h+=X[i]+K[i]
'&mov ($a2,$a)',
- '&ror ($a1,$Sigma0[1]-$Sigma0[0])',
'&xor ($a4,$g)', # Ch(e,f,g)=((f^g)&e)^g
+ '&ror ($a1,$Sigma0[1]-$Sigma0[0])',
'&xor ($a2,$b)', # a^b, b^c in next round
- '&ror ($a0,$Sigma1[0])', # Sigma1(e)
'&add ($h,$a4)', # h+=Ch(e,f,g)
+ '&ror ($a0,$Sigma1[0])', # Sigma1(e)
'&and ($a3,$a2)', # (b^c)&(a^b)
'&xor ($a1,$a)',
'&add ($h,$a0)', # h+=Sigma1(e)
'&xor ($a3,$b)', # Maj(a,b,c)=Ch(a^b,c,b)
- '&add ($d,$h)', # d+=h
'&ror ($a1,$Sigma0[0])', # Sigma0(a)
+ '&add ($d,$h)', # d+=h
'&add ($h,$a3)', # h+=Maj(a,b,c)
'&mov ($a0,$d)',
___
$code.=<<___;
- movdqa $TABLE+`$SZ*2*$rounds`+32(%rip),$t4
- movdqa $TABLE+`$SZ*2*$rounds`+64(%rip),$t5
+ #movdqa $TABLE+`$SZ*2*$rounds`+32(%rip),$t4
+ #movdqa $TABLE+`$SZ*2*$rounds`+64(%rip),$t5
jmp .Lloop_ssse3
.align 16
.Lloop_ssse3:
eval(shift(@insns));
eval(shift(@insns));
}
- } else { # squeeze extra 3% on Westmere and Atom
+ } else { # squeeze extra 4% on Westmere and 19% on Atom
eval(shift(@insns)); #@
- eval(shift(@insns));
&movdqa ($t0,@X[1]);
eval(shift(@insns));
+ eval(shift(@insns));
&movdqa ($t3,@X[3]);
+ eval(shift(@insns)); #@
+ eval(shift(@insns));
eval(shift(@insns));
eval(shift(@insns)); #@
eval(shift(@insns));
&palignr ($t0,@X[0],$SZ); # X[1..4]
- eval(shift(@insns)); #@
eval(shift(@insns));
- &palignr ($t3,@X[2],$SZ); # X[9..12]
eval(shift(@insns));
+ &palignr ($t3,@X[2],$SZ); # X[9..12]
eval(shift(@insns));
eval(shift(@insns));
eval(shift(@insns));
eval(shift(@insns)); #@
- eval(shift(@insns));
&movdqa ($t1,$t0);
eval(shift(@insns));
+ eval(shift(@insns));
&movdqa ($t2,$t0);
eval(shift(@insns)); #@
eval(shift(@insns));
- eval(shift(@insns));
&psrld ($t0,$sigma0[2]);
eval(shift(@insns));
eval(shift(@insns));
eval(shift(@insns));
&paddd (@X[0],$t3); # X[0..3] += X[9..12]
- eval(shift(@insns));
eval(shift(@insns)); #@
eval(shift(@insns));
&psrld ($t2,$sigma0[0]);
eval(shift(@insns));
eval(shift(@insns));
- eval(shift(@insns)); #@
- eval(shift(@insns));
&pshufd ($t3,@X[3],0b11111010); # X[4..15]
eval(shift(@insns));
+ eval(shift(@insns)); #@
&pslld ($t1,8*$SZ-$sigma0[1]);
eval(shift(@insns));
+ eval(shift(@insns));
&pxor ($t0,$t2);
eval(shift(@insns)); #@
eval(shift(@insns));
- &psrld ($t2,$sigma0[1]-$sigma0[0]);
+ eval(shift(@insns));
eval(shift(@insns)); #@
+ &psrld ($t2,$sigma0[1]-$sigma0[0]);
eval(shift(@insns));
&pxor ($t0,$t1);
eval(shift(@insns));
eval(shift(@insns));
&pslld ($t1,$sigma0[1]-$sigma0[0]);
eval(shift(@insns));
+ eval(shift(@insns));
&pxor ($t0,$t2);
eval(shift(@insns));
eval(shift(@insns)); #@
- eval(shift(@insns));
&movdqa ($t2,$t3);
eval(shift(@insns));
- eval(shift(@insns)); #@
eval(shift(@insns));
&pxor ($t0,$t1); # sigma0(X[1..4])
+ eval(shift(@insns)); #@
eval(shift(@insns));
eval(shift(@insns));
&psrld ($t3,$sigma1[2]);
eval(shift(@insns));
eval(shift(@insns));
&paddd (@X[0],$t0); # X[0..3] += sigma0(X[1..4])
- eval(shift(@insns));
eval(shift(@insns)); #@
eval(shift(@insns));
- eval(shift(@insns));
&psrlq ($t2,$sigma1[0]);
eval(shift(@insns));
- eval(shift(@insns)); #@
eval(shift(@insns));
eval(shift(@insns));
&pxor ($t3,$t2);
+ eval(shift(@insns)); #@
+ eval(shift(@insns));
eval(shift(@insns));
eval(shift(@insns)); #@
&psrlq ($t2,$sigma1[1]-$sigma1[0]);
eval(shift(@insns));
- eval(shift(@insns)); #@
eval(shift(@insns));
&pxor ($t3,$t2);
+ eval(shift(@insns)); #@
eval(shift(@insns));
eval(shift(@insns));
+ #&pshufb ($t3,$t4); # sigma1(X[14..15])
+ &pshufd ($t3,$t3,0b10000000);
eval(shift(@insns));
- &pshufb ($t3,$t4); # sigma1(X[14..15])
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &psrldq ($t3,8);
eval(shift(@insns));
eval(shift(@insns)); #@
eval(shift(@insns));
eval(shift(@insns)); #@
&paddd (@X[0],$t3); # X[0..1] += sigma1(X[14..15])
eval(shift(@insns));
- &pshufd ($t3,@X[0],0b01010000); # X[16..17]
eval(shift(@insns));
eval(shift(@insns));
+ &pshufd ($t3,@X[0],0b01010000); # X[16..17]
eval(shift(@insns));
+ eval(shift(@insns)); #@
eval(shift(@insns));
&movdqa ($t2,$t3);
eval(shift(@insns));
- eval(shift(@insns)); #@
eval(shift(@insns));
&psrld ($t3,$sigma1[2]);
eval(shift(@insns));
- &psrlq ($t2,$sigma1[0]);
- eval(shift(@insns));
eval(shift(@insns)); #@
+ &psrlq ($t2,$sigma1[0]);
eval(shift(@insns));
eval(shift(@insns));
&pxor ($t3,$t2);
+ eval(shift(@insns)); #@
+ eval(shift(@insns));
eval(shift(@insns));
eval(shift(@insns)); #@
eval(shift(@insns));
&psrlq ($t2,$sigma1[1]-$sigma1[0]);
- eval(shift(@insns)); #@
eval(shift(@insns));
eval(shift(@insns));
eval(shift(@insns));
&pxor ($t3,$t2);
eval(shift(@insns));
eval(shift(@insns));
- &movdqa ($t2,16*2*$j."($Tbl)");
eval(shift(@insns)); #@
+ #&pshufb ($t3,$t5);
+ &pshufd ($t3,$t3,0b00001000);
eval(shift(@insns));
- &pshufb ($t3,$t5);
eval(shift(@insns));
+ &movdqa ($t2,16*2*$j."($Tbl)");
eval(shift(@insns)); #@
eval(shift(@insns));
+ &pslldq ($t3,8);
eval(shift(@insns));
eval(shift(@insns));
- &paddd (@X[0],$t3); # X[2..3] += sigma1(X[16..17])
eval(shift(@insns));
+ &paddd (@X[0],$t3); # X[2..3] += sigma1(X[16..17])
+ eval(shift(@insns)); #@
eval(shift(@insns));
eval(shift(@insns));
}
$code.=<<___;
.Lprologue_xop:
- vzeroall
+ vzeroupper
mov $SZ*0($ctx),$A
mov $SZ*1($ctx),$B
mov $SZ*2($ctx),$C
jb .Lloop_xop
mov $_rsp,%rsi
- vzeroall
+ vzeroupper
___
$code.=<<___ if ($win64);
movaps 16*$SZ+32(%rsp),%xmm6
$code.=<<___;
.Lprologue_avx:
- vzeroall
+ vzeroupper
mov $SZ*0($ctx),$A
mov $SZ*1($ctx),$B
mov $SZ*2($ctx),$C
jb .Lloop_avx
mov $_rsp,%rsi
- vzeroall
+ vzeroupper
___
$code.=<<___ if ($win64);
movaps 16*$SZ+32(%rsp),%xmm6
$code.=<<___;
.Lprologue_avx2:
- vzeroall
+ vzeroupper
sub \$-16*$SZ,$inp # inp++, size optimization
mov $SZ*0($ctx),$A
- xor %r12,%r12 # borrow $T1
+ mov $inp,%r12 # borrow $T1
mov $SZ*1($ctx),$B
cmp %rdx,$inp # $_end
mov $SZ*2($ctx),$C
- sete %r12b
+ cmove %rsp,%r12 # next block or random data
mov $SZ*3($ctx),$D
mov $SZ*4($ctx),$E
mov $SZ*5($ctx),$F
jmp .Loop_avx2
.align 16
.Loop_avx2:
- shl \$`log(16*$SZ)/log(2)`,%r12
vmovdqa $TABLE+`$SZ*2*$rounds`(%rip),$t3
- neg %r12
- vmovdqu -16*$SZ+0($inp),$t0
- add $inp,%r12
- vmovdqu -16*$SZ+32($inp),$t1
- vmovdqu (%r12),@X[2] # next or same input block
- vmovdqu 32(%r12),@X[3]
-
- vperm2i128 \$0x20,@X[2],$t0,@X[0]
+ vmovdqu -16*$SZ+0($inp),%xmm0
+ vmovdqu -16*$SZ+16($inp),%xmm1
+ vmovdqu -16*$SZ+32($inp),%xmm2
+ vmovdqu -16*$SZ+48($inp),%xmm3
#mov $inp,$_inp # offload $inp
- vperm2i128 \$0x31,@X[2],$t0,@X[1]
- vperm2i128 \$0x20,@X[3],$t1,@X[2]
- vperm2i128 \$0x31,@X[3],$t1,@X[3]
+ vinserti128 \$1,(%r12),@X[0],@X[0]
+ vinserti128 \$1,16(%r12),@X[1],@X[1]
+ vpshufb $t3,@X[0],@X[0]
+ vinserti128 \$1,32(%r12),@X[2],@X[2]
+ vpshufb $t3,@X[1],@X[1]
+ vinserti128 \$1,48(%r12),@X[3],@X[3]
lea $TABLE(%rip),$Tbl
- vpshufb $t3,@X[0],@X[0]
- vpshufb $t3,@X[1],@X[1]
vpshufb $t3,@X[2],@X[2]
vpaddd 0x00($Tbl),@X[0],$t0
vpshufb $t3,@X[3],@X[3]
jmp .Loop_avx2
.align 16
.Loop_avx2:
- shl \$`log(16*$SZ)/log(2)`,%r12
- vmovdqu -16*$SZ($inp),$t0
- neg %r12
- vmovdqu -16*$SZ+32($inp),$t1
- add $inp,%r12
- vmovdqu -16*$SZ+64($inp),$t2
- vmovdqu -16*$SZ+96($inp),$t3
- vmovdqu (%r12),@X[4] # next or same block
- vmovdqu 32(%r12),@X[5]
- vmovdqu 64(%r12),@X[6]
- vmovdqu 96(%r12),@X[7]
-
- vperm2i128 \$0x20,@X[4],$t0,@X[0]
- #mov $inp,$_inp # offload $inp
- vperm2i128 \$0x31,@X[4],$t0,@X[1]
- vperm2i128 \$0x20,@X[5],$t1,@X[2]
- vperm2i128 \$0x31,@X[5],$t1,@X[3]
- vperm2i128 \$0x20,@X[6],$t2,@X[4]
- vperm2i128 \$0x31,@X[6],$t2,@X[5]
- vmovdqa $TABLE+`$SZ*2*$rounds`(%rip),$t2
- vperm2i128 \$0x20,@X[7],$t3,@X[6]
- vperm2i128 \$0x31,@X[7],$t3,@X[7]
-
- vpshufb $t2,@X[0],@X[0]
+ vmovdqu -16*$SZ($inp),%xmm0
+ vmovdqu -16*$SZ+16($inp),%xmm1
+ vmovdqu -16*$SZ+32($inp),%xmm2
lea $TABLE+0x80(%rip),$Tbl # size optimization
- vpshufb $t2,@X[1],@X[1]
- vpshufb $t2,@X[2],@X[2]
- vpshufb $t2,@X[3],@X[3]
- vpshufb $t2,@X[4],@X[4]
- vpshufb $t2,@X[5],@X[5]
+ vmovdqu -16*$SZ+48($inp),%xmm3
+ vmovdqu -16*$SZ+64($inp),%xmm4
+ vmovdqu -16*$SZ+80($inp),%xmm5
+ vmovdqu -16*$SZ+96($inp),%xmm6
+ vmovdqu -16*$SZ+112($inp),%xmm7
+ #mov $inp,$_inp # offload $inp
+ vmovdqa `$SZ*2*$rounds-0x80`($Tbl),$t2
+ vinserti128 \$1,(%r12),@X[0],@X[0]
+ vinserti128 \$1,16(%r12),@X[1],@X[1]
+ vpshufb $t2,@X[0],@X[0]
+ vinserti128 \$1,32(%r12),@X[2],@X[2]
+ vpshufb $t2,@X[1],@X[1]
+ vinserti128 \$1,48(%r12),@X[3],@X[3]
+ vpshufb $t2,@X[2],@X[2]
+ vinserti128 \$1,64(%r12),@X[4],@X[4]
+ vpshufb $t2,@X[3],@X[3]
+ vinserti128 \$1,80(%r12),@X[5],@X[5]
+ vpshufb $t2,@X[4],@X[4]
+ vinserti128 \$1,96(%r12),@X[6],@X[6]
+ vpshufb $t2,@X[5],@X[5]
+ vinserti128 \$1,112(%r12),@X[7],@X[7]
+
vpaddq -0x80($Tbl),@X[0],$t0
vpshufb $t2,@X[6],@X[6]
vpaddq -0x60($Tbl),@X[1],$t1
add $SZ*5($ctx),$F
lea `2*16*$SZ`($inp),$inp # inp+=2
add $SZ*6($ctx),$G
- xor %r12,%r12
+ mov $inp,%r12
add $SZ*7($ctx),$H
cmp $_end,$inp
mov $A,$SZ*0($ctx)
+ cmove %rsp,%r12 # next block or stale data
mov $B,$SZ*1($ctx)
mov $C,$SZ*2($ctx)
mov $D,$SZ*3($ctx)
mov $G,$SZ*6($ctx)
mov $H,$SZ*7($ctx)
- sete %r12b
jbe .Loop_avx2
lea (%rsp),$Tbl
.Ldone_avx2:
lea ($Tbl),%rsp
mov $_rsp,%rsi
- vzeroall
+ vzeroupper
___
$code.=<<___ if ($win64);
movaps 16*$SZ+32(%rsp),%xmm6