From 07e29c1234177899aceec350cf6db3dbdc60beef Mon Sep 17 00:00:00 2001 From: Andy Polyakov Date: Sun, 23 May 2010 12:37:01 +0000 Subject: [PATCH] ghash-x86.pl: MMX optimization (+20-40%) and commentary update. --- crypto/modes/asm/ghash-x86.pl | 171 ++++++++++++++++++---------------- 1 file changed, 89 insertions(+), 82 deletions(-) diff --git a/crypto/modes/asm/ghash-x86.pl b/crypto/modes/asm/ghash-x86.pl index c21664e53c..d74c33713a 100644 --- a/crypto/modes/asm/ghash-x86.pl +++ b/crypto/modes/asm/ghash-x86.pl @@ -7,7 +7,7 @@ # details see http://www.openssl.org/~appro/cryptogams/. # ==================================================================== # -# March 2010 +# March, May 2010 # # The module implements "4-bit" GCM GHASH function and underlying # single multiplication operation in GF(2^128). "4-bit" means that it @@ -20,10 +20,10 @@ # gcc 2.95.3(*) MMX assembler x86 assembler # # Pentium 100/112(**) - 50 -# PIII 63 /77 16 24 -# P4 96 /122 30 84(***) -# Opteron 50 /71 21 30 -# Core2 54 /68 12.5 18 +# PIII 63 /77 14.5 24 +# P4 96 /122 24.5 84(***) +# Opteron 50 /71 14.5 30 +# Core2 54 /68 10.5 18 # # (*) gcc 3.4.x was observed to generate few percent slower code, # which is one of reasons why 2.95.3 results were chosen, @@ -33,9 +33,10 @@ # position-independent; # (***) see comment in non-MMX routine for further details; # -# To summarize, it's >2-3 times faster than gcc-generated code. To +# To summarize, it's >2-4 times faster than gcc-generated code. To # anchor it to something else SHA1 assembler processes one byte in -# 11-13 cycles on contemporary x86 cores. +# 11-13 cycles on contemporary x86 cores. As for choice of MMX in +# particular, see comment at the end of the file... # May 2010 # @@ -317,17 +318,24 @@ if (!$x86only) {{{ &static_label("rem_4bit"); -sub mmx_loop() { -# MMX version performs 2.8 times better on P4 (see comment in non-MMX -# routine for further details), 40% better on Opteron and Core2, 50% -# better on PIII... In other words effort is considered to be well -# spent... - my $inp = shift; - my $rem_4bit = shift; - my $cnt = $Zhh; +&function_begin_B("_mmx_gmult_4bit_inner"); +# MMX version performs 3.5 times better on P4 (see comment in non-MMX +# routine for further details), 100% better on Opteron, ~70% better +# on Core2 and PIII... In other words effort is considered to be well +# spent... Since initial release the loop was unrolled in order to +# "liberate" register previously used as loop counter. Instead it's +# used to optimize critical path in 'Z.hi ^= rem_4bit[Z.lo&0xf]'. +# The path involves move of Z.lo from MMX to integer register, +# effective address calculation and finally merge of value to Z.hi. +# Reference to rem_4bit is scheduled so late that I had to >>4 +# rem_4bit elements. This resulted in 20-45% procent improvement +# on contemporary ยต-archs. +{ + my $cnt; + my $rem_4bit = "eax"; + my @rem = ($Zhh,$Zll); my $nhi = $Zhl; my $nlo = $Zlh; - my $rem = $Zll; my ($Zlo,$Zhi) = ("mm0","mm1"); my $tmp = "mm2"; @@ -335,80 +343,52 @@ sub mmx_loop() { &xor ($nlo,$nlo); # avoid partial register stalls on PIII &mov ($nhi,$Zll); &mov (&LB($nlo),&LB($nhi)); - &mov ($cnt,14); &shl (&LB($nlo),4); &and ($nhi,0xf0); &movq ($Zlo,&QWP(8,$Htbl,$nlo)); &movq ($Zhi,&QWP(0,$Htbl,$nlo)); - &movd ($rem,$Zlo); - &jmp (&label("mmx_loop")); - - &set_label("mmx_loop",16); - &psrlq ($Zlo,4); - &and ($rem,0xf); - &movq ($tmp,$Zhi); - &psrlq ($Zhi,4); - &pxor ($Zlo,&QWP(8,$Htbl,$nhi)); - &mov (&LB($nlo),&BP(0,$inp,$cnt)); - &psllq ($tmp,60); - &pxor ($Zhi,&QWP(0,$rem_4bit,$rem,8)); - &dec ($cnt); - &movd ($rem,$Zlo); - &pxor ($Zhi,&QWP(0,$Htbl,$nhi)); - &mov ($nhi,$nlo); - &pxor ($Zlo,$tmp); - &js (&label("mmx_break")); + &movd ($rem[0],$Zlo); + + for ($cnt=28;$cnt>=-2;$cnt--) { + my $odd = $cnt&1; + my $nix = $odd ? $nlo : $nhi; + + &shl (&LB($nlo),4) if ($odd); + &psrlq ($Zlo,4); + &movq ($tmp,$Zhi); + &psrlq ($Zhi,4); + &pxor ($Zlo,&QWP(8,$Htbl,$nix)); + &mov (&LB($nlo),&BP($cnt/2,$inp)) if (!$odd && $cnt>=0); + &psllq ($tmp,60); + &and ($nhi,0xf0) if ($odd); + &pxor ($Zhi,&QWP(0,$rem_4bit,$rem[1],8)) if ($cnt<28); + &and ($rem[0],0xf); + &pxor ($Zhi,&QWP(0,$Htbl,$nix)); + &mov ($nhi,$nlo) if (!$odd && $cnt>=0); + &movd ($rem[1],$Zlo); + &pxor ($Zlo,$tmp); + + push (@rem,shift(@rem)); # "rotate" registers + } - &shl (&LB($nlo),4); - &and ($rem,0xf); - &psrlq ($Zlo,4); - &and ($nhi,0xf0); - &movq ($tmp,$Zhi); - &psrlq ($Zhi,4); - &pxor ($Zlo,&QWP(8,$Htbl,$nlo)); - &psllq ($tmp,60); - &pxor ($Zhi,&QWP(0,$rem_4bit,$rem,8)); - &movd ($rem,$Zlo); - &pxor ($Zhi,&QWP(0,$Htbl,$nlo)); - &pxor ($Zlo,$tmp); - &jmp (&label("mmx_loop")); - - &set_label("mmx_break",16); - &shl (&LB($nlo),4); - &and ($rem,0xf); - &psrlq ($Zlo,4); - &and ($nhi,0xf0); - &movq ($tmp,$Zhi); - &psrlq ($Zhi,4); - &pxor ($Zlo,&QWP(8,$Htbl,$nlo)); - &psllq ($tmp,60); - &pxor ($Zhi,&QWP(0,$rem_4bit,$rem,8)); - &movd ($rem,$Zlo); - &pxor ($Zhi,&QWP(0,$Htbl,$nlo)); - &pxor ($Zlo,$tmp); - - &psrlq ($Zlo,4); - &and ($rem,0xf); - &movq ($tmp,$Zhi); - &psrlq ($Zhi,4); - &pxor ($Zlo,&QWP(8,$Htbl,$nhi)); - &psllq ($tmp,60); - &pxor ($Zhi,&QWP(0,$rem_4bit,$rem,8)); - &movd ($rem,$Zlo); - &pxor ($Zhi,&QWP(0,$Htbl,$nhi)); - &pxor ($Zlo,$tmp); + &mov ($inp,&DWP(4,$rem_4bit,$rem[1],8)); # last rem_4bit[rem] &psrlq ($Zlo,32); # lower part of Zlo is already there &movd ($Zhl,$Zhi); &psrlq ($Zhi,32); &movd ($Zlh,$Zlo); &movd ($Zhh,$Zhi); + &shl ($inp,4); # compensate for rem_4bit[i] being >>4 &bswap ($Zll); &bswap ($Zhl); &bswap ($Zlh); + &xor ($Zhh,$inp); &bswap ($Zhh); + + &ret (); } +&function_end_B("_mmx_gmult_4bit_inner"); &function_begin("gcm_gmult_4bit_mmx"); &mov ($inp,&wparam(0)); # load Xi @@ -421,8 +401,9 @@ sub mmx_loop() { &movz ($Zll,&BP(15,$inp)); - &mmx_loop($inp,"eax"); + &call ("_mmx_gmult_4bit_inner"); + &mov ($inp,&wparam(0)); # load Xi &emms (); &mov (&DWP(12,$inp),$Zll); &mov (&DWP(4,$inp),$Zhl); @@ -458,15 +439,18 @@ sub mmx_loop() { &xor ($Zhl,&DWP(4,$inp)); &xor ($Zlh,&DWP(8,$inp)); &xor ($Zhh,&DWP(0,$inp)); + &mov (&wparam(2),$inp); &mov (&DWP(12,"esp"),$Zll); &mov (&DWP(4,"esp"),$Zhl); &mov (&DWP(8,"esp"),$Zlh); &mov (&DWP(0,"esp"),$Zhh); + &mov ($inp,"esp"); &shr ($Zll,24); - &mmx_loop("esp","eax"); + &call ("_mmx_gmult_4bit_inner"); + &mov ($inp,&wparam(2)); &lea ($inp,&DWP(16,$inp)); &cmp ($inp,&wparam(3)); &jb (&label("mmx_outer_loop")); @@ -552,8 +536,8 @@ if (1) { # Algorithm 9 with <<1 twist. # candidate for interleaving with 64x64 # multiplication. Pre-modulo-scheduled loop # was found to be ~20% faster than Algorithm 5 - # below. Algorithm 9 was then chosen and - # optimized further... + # below. Algorithm 9 was therefore chosen for + # further optimization... sub reduction_alg9 { # 17/13 times faster than Intel version my ($Xhi,$Xi) = @_; @@ -567,7 +551,7 @@ my ($Xhi,$Xi) = @_; &psllq ($Xi,57); # &movdqa ($T2,$Xi); # &pslldq ($Xi,8); - &psrldq ($T2,8); # + &psrldq ($T2,8); # &pxor ($Xi,$T1); &pxor ($Xhi,$T2); # @@ -952,11 +936,34 @@ my ($Xhi,$Xi)=@_; }} # $sse2 &set_label("rem_4bit",64); - &data_word(0,0x0000<<16,0,0x1C20<<16,0,0x3840<<16,0,0x2460<<16); - &data_word(0,0x7080<<16,0,0x6CA0<<16,0,0x48C0<<16,0,0x54E0<<16); - &data_word(0,0xE100<<16,0,0xFD20<<16,0,0xD940<<16,0,0xC560<<16); - &data_word(0,0x9180<<16,0,0x8DA0<<16,0,0xA9C0<<16,0,0xB5E0<<16); + &data_word(0,0x0000<<12,0,0x1C20<<12,0,0x3840<<12,0,0x2460<<12); + &data_word(0,0x7080<<12,0,0x6CA0<<12,0,0x48C0<<12,0,0x54E0<<12); + &data_word(0,0xE100<<12,0,0xFD20<<12,0,0xD940<<12,0,0xC560<<12); + &data_word(0,0x9180<<12,0,0x8DA0<<12,0,0xA9C0<<12,0,0xB5E0<<12); }}} # !$x86only &asciz("GHASH for x86, CRYPTOGAMS by "); &asm_finish(); + +# A question was risen about choice of vanilla MMX. Or rather why wasn't +# SSE2 chosen instead? In addition to the fact that MMX runs on legacy +# CPUs such as PIII, "4-bit" MMX version was observed to provide better +# performance than *corresponding* SSE2 one even on contemporary CPUs. +# SSE2 results were provided by Peter-Michael Hager. He maintains SSE2 +# implementation featuring full range of lookup-table sizes, but with +# per-invocation lookup table setup. Latter means that table size is +# chosen depending on how much data is to be hashed in every given call, +# more data - larger table. Best reported result for Core2 is ~4 cycles +# per processed byte out of 64KB block. Recall that this number accounts +# even for 64KB table setup overhead. As discussed in gcm128.c we choose +# to be more conservative in respect to lookup table sizes, but how +# do the results compare? As per table in the beginning, minimalistic +# MMX version delivers ~11 cycles on same platform. As also discussed in +# gcm128.c, next in line "8-bit Shoup's" method should deliver twice the +# performance of "4-bit" one. It should be also be noted that in SSE2 +# case improvement can be "super-linear," i.e. more than twice, mostly +# because >>8 maps to single instruction on SSE2 register. This is +# unlike "4-bit" case when >>4 maps to same amount of instructions in +# both MMX and SSE2 cases. Bottom line is that switch to SSE2 is +# considered to be justifiable only in case we choose to implement +# "8-bit" method... -- 2.25.1