wbrawner/openssl
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
openssl/crypto/sha/asm/keccak1600-avx2.pl
Andy Polyakov d212b98b36 sha/asm/keccak1600-avx2.pl: optimized remodelled version. 
New register usage pattern allows to achieve sligtly better
performance. Not as much as I hoped for. Performance is believed
to be limited by irreconcilable write-back conflicts, rather than
lack of computational resources or data dependencies.

Reviewed-by: Rich Salz <rsalz@openssl.org>
2017-07-15 23:04:38 +02:00

480 lines
16 KiB
Raku
Executable file
Raw Blame History

#!/usr/bin/env perl
# Copyright 2017 The OpenSSL Project Authors. All Rights Reserved.
#
# Licensed under the OpenSSL license (the "License"). You may not use
# this file except in compliance with the License. You can obtain a copy
# in the file LICENSE in the source distribution or at
# https://www.openssl.org/source/license.html
#
# ====================================================================
# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
# project. The module is, however, dual licensed under OpenSSL and
# CRYPTOGAMS licenses depending on where you obtain it. For further
# details see http://www.openssl.org/~appro/cryptogams/.
# ====================================================================
#
# Keccak-1600 for AVX2.
#
# July 2017.
#
# To paraphrase Gilles Van Assche, if you contemplate Fig. 2.3 on page
# 20 of The Keccak reference [or Fig. 5 of FIPS PUB 202], and load data
# other than A[0][0] in magic order into 6 [256-bit] registers, *each
# dedicated to one axis*, Pi permutation is reduced to intra-register
# shuffles...
#
# It makes other steps more intricate, but overall, is it a win? To be
# more specific index permutations organized by quadruples are:
#
# [4][4] [3][3] [2][2] [1][1]<-+
# [0][4] [0][3] [0][2] [0][1]<-+
# [3][0] [1][0] [4][0] [2][0] |
# [4][3] [3][1] [2][4] [1][2] |
# [3][4] [1][3] [4][2] [2][1] |
# [2][3] [4][1] [1][4] [3][2] |
# [2][2] [4][4] [1][1] [3][3] -+
#
# This however is highly impractical for Theta and Chi. What would help
# Theta is if x indices were aligned column-wise, or in other words:
#
# [0][4] [0][3] [0][2] [0][1]
# [3][0] [1][0] [4][0] [2][0]
#vpermq([4][3] [3][1] [2][4] [1][2], 0b01110010)
# [2][4] [4][3] [1][2] [3][1]
#vpermq([4][2] [3][4] [2][1] [1][3], 0b10001101)
# [3][4] [1][3] [4][2] [2][1]
#vpermq([2][3] [4][1] [1][4] [3][2], 0b01110010)
# [1][4] [2][3] [3][2] [4][1]
#vpermq([1][1] [2][2] [3][3] [4][4], 0b00011011)
# [4][4] [3][3] [2][2] [1][1]
#
# So here we have it, lines not marked with vpermq() represent the magic
# order in which data is to be loaded and maintained. [And lines marked
# with vpermq() represent Pi circular permutation in chosen layout. Note
# that first step is permutation-free.] A[0][0] is loaded to register of
# its own, to all lanes. [A[0][0] is not part of Pi permutation or Rho.]
# Digits in variables' names denote right-most coordinates:
my ($A00, # [0][0] [0][0] [0][0] [0][0] # %ymm0
$A01, # [0][4] [0][3] [0][2] [0][1] # %ymm1
$A20, # [3][0] [1][0] [4][0] [2][0] # %ymm2
$A31, # [2][4] [4][3] [1][2] [3][1] # %ymm3
$A21, # [3][4] [1][3] [4][2] [2][1] # %ymm4
$A41, # [1][4] [2][3] [3][2] [4][1] # %ymm5
$A11) = # [4][4] [3][3] [2][2] [1][1] # %ymm6
map("%ymm$_",(0..6));
# We also need to map the magic order into offsets within structure:
my @A_jagged = ([0,0], [1,0], [1,1], [1,2], [1,3], # [0][0..4]
[2,2], [6,0], [3,1], [4,2], [5,3], # [1][0..4]
[2,0], [4,0], [6,1], [5,2], [3,3], # [2][0..4]
[2,3], [3,0], [5,1], [6,2], [4,3], # [3][0..4]
[2,1], [5,0], [4,1], [3,2], [6,3]); # [4][0..4]
@A_jagged = map(8*($$_[0]*4+$$_[1]), @A_jagged); # ... and now linear
# But on the other hand Chi is much better off if y indices were aligned
# column-wise, not x. For this reason we have to shuffle data prior
# Chi and revert it afterwards. Prior shuffle is naturally merged with
# Pi itself:
#
# [0][4] [0][3] [0][2] [0][1]
# [3][0] [1][0] [4][0] [2][0]
#vpermq([4][3] [3][1] [2][4] [1][2], 0b01110010)
#vpermq([2][4] [4][3] [1][2] [3][1], 0b00011011) = 0b10001101
# [3][1] [1][2] [4][3] [2][4]
#vpermq([4][2] [3][4] [2][1] [1][3], 0b10001101)
#vpermq([3][4] [1][3] [4][2] [2][1], 0b11100100) = 0b10001101
# [3][4] [1][3] [4][2] [2][1]
#vpermq([2][3] [4][1] [1][4] [3][2], 0b01110010)
#vpermq([1][4] [2][3] [3][2] [4][1], 0b01110010) = 0b00011011
# [3][2] [1][4] [4][1] [2][3]
#vpermq([1][1] [2][2] [3][3] [4][4], 0b00011011)
#vpermq([4][4] [3][3] [2][2] [1][1], 0b10001101) = 0b01110010
# [3][3] [1][1] [4][4] [2][2]
#
# And reverse post-Chi permutation:
#
# [0][4] [0][3] [0][2] [0][1]
# [3][0] [1][0] [4][0] [2][0]
#vpermq([3][1] [1][2] [4][3] [2][4], 0b00011011)
# [2][4] [4][3] [1][2] [3][1]
#vpermq([3][4] [1][3] [4][2] [2][1], 0b11100100) = nop :-)
# [3][4] [1][3] [4][2] [2][1]
#vpermq([3][2] [1][4] [4][1] [2][3], 0b10001101)
# [1][4] [2][3] [3][2] [4][1]
#vpermq([3][3] [1][1] [4][4] [2][2], 0b01110010)
# [4][4] [3][3] [2][2] [1][1]
#
########################################################################
# Numbers are cycles per processed byte out of large message.
#
# r=1088(*)
#
# Haswell 8.7/+10%
# Skylake 7.8/+20%
# Ryzen 17(**)
#
# (*) Corresponds to SHA3-256. Percentage after slash is improvement
# coefficient in comparison to scalar keccak1600-x86_64.pl.
# (**) It's expected that Ryzen performs poorly, because instruction
# issue rate is limited to two AVX2 instructions per cycle and
# in addition vpblendd is reportedly bound to specific port.
# Obviously this code path should not be executed on Ryzen.
my @T = map("%ymm$_",(7..15));
my ($C14,$C00,$D00,$D14) = @T[5..8];
$code.=<<___;
.text
.type __KeccakF1600,\@function
.align 32
__KeccakF1600:
lea rhotates_left+96(%rip),%r8
lea rhotates_right+96(%rip),%r9
lea iotas(%rip),%r10
mov \$24,%eax
jmp .Loop_avx2
.align 32
.Loop_avx2:
######################################### Theta
vpshufd \$0b01001110,$A20,$C00
vpxor $A31,$A41,$C14
vpxor $A11,$A21,@T[2]
vpxor $A01,$C14,$C14
vpxor @T[2],$C14,$C14 # C[1..4]
vpermq \$0b10010011,$C14,@T[4]
vpxor $A20,$C00,$C00
vpermq \$0b01001110,$C00,@T[0]
vpsrlq \$63,$C14,@T[1]
vpaddq $C14,$C14,@T[2]
vpor @T[2],@T[1],@T[1] # ROL64(C[1..4],1)
vpermq \$0b00111001,@T[1],$D14
vpxor @T[4],@T[1],$D00
vpermq \$0b00000000,$D00,$D00 # D[0..0] = ROL64(C[1],1) ^ C[4]
vpxor $A00,$C00,$C00
vpxor @T[0],$C00,$C00 # C[0..0]
vpsrlq \$63,$C00,@T[0]
vpaddq $C00,$C00,@T[1]
vpor @T[0],@T[1],@T[1] # ROL64(C[0..0],1)
vpxor $D00,$A20,$A20 # ^= D[0..0]
vpxor $D00,$A00,$A00 # ^= D[0..0]
vpblendd \$0b11000000,@T[1],$D14,$D14
vpblendd \$0b00000011,$C00,@T[4],@T[4]
vpxor @T[4],$D14,$D14 # D[1..4] = ROL64(C[2..4,0),1) ^ C[0..3]
######################################### Rho + Pi + pre-Chi shuffle
vpsllvq 0*32-96(%r8),$A20,@T[3]
vpsrlvq 0*32-96(%r9),$A20,$A20
vpor @T[3],$A20,$A20
vpxor $D14,$A31,$A31 # ^= D[1..4] from Theta
vpsllvq 2*32-96(%r8),$A31,@T[4]
vpsrlvq 2*32-96(%r9),$A31,$A31
vpor @T[4],$A31,$A31
vpxor $D14,$A21,$A21 # ^= D[1..4] from Theta
vpsllvq 3*32-96(%r8),$A21,@T[5]
vpsrlvq 3*32-96(%r9),$A21,$A21
vpor @T[5],$A21,$A21
vpxor $D14,$A41,$A41 # ^= D[1..4] from Theta
vpsllvq 4*32-96(%r8),$A41,@T[6]
vpsrlvq 4*32-96(%r9),$A41,$A41
vpor @T[6],$A41,$A41
vpxor $D14,$A11,$A11 # ^= D[1..4] from Theta
vpermq \$0b10001101,$A20,@T[3] # $A20 -> future $A31
vpermq \$0b10001101,$A31,@T[4] # $A31 -> future $A21
vpsllvq 5*32-96(%r8),$A11,@T[7]
vpsrlvq 5*32-96(%r9),$A11,@T[1]
vpor @T[7],@T[1],@T[1] # $A11 -> future $A01
vpxor $D14,$A01,$A01 # ^= D[1..4] from Theta
vpermq \$0b00011011,$A21,@T[5] # $A21 -> future $A41
vpermq \$0b01110010,$A41,@T[6] # $A41 -> future $A11
vpsllvq 1*32-96(%r8),$A01,@T[8]
vpsrlvq 1*32-96(%r9),$A01,@T[2]
vpor @T[8],@T[2],@T[2] # $A01 -> future $A20
######################################### Chi
vpsrldq \$8,@T[1],@T[7]
vpandn @T[7],@T[1],@T[0] # tgting [0][0] [0][0] [0][0] [0][0]
vpblendd \$0b00001100,@T[6],@T[2],$A31 # [4][4] [2][0]
vpblendd \$0b00001100,@T[2],@T[4],@T[8] # [4][0] [2][1]
vpblendd \$0b00001100,@T[4],@T[3],$A41 # [4][2] [2][4]
vpblendd \$0b00001100,@T[3],@T[2],@T[7] # [4][3] [2][0]
vpblendd \$0b00110000,@T[4],$A31,$A31 # [1][3] [4][4] [2][0]
vpblendd \$0b00110000,@T[5],@T[8],@T[8] # [1][4] [4][0] [2][1]
vpblendd \$0b00110000,@T[2],$A41,$A41 # [1][0] [4][2] [2][4]
vpblendd \$0b00110000,@T[6],@T[7],@T[7] # [1][1] [4][3] [2][0]
vpblendd \$0b11000000,@T[5],$A31,$A31 # [3][2] [1][3] [4][4] [2][0]
vpblendd \$0b11000000,@T[6],@T[8],@T[8] # [3][3] [1][4] [4][0] [2][1]
vpblendd \$0b11000000,@T[6],$A41,$A41 # [3][3] [1][0] [4][2] [2][4]
vpblendd \$0b11000000,@T[4],@T[7],@T[7] # [3][4] [1][1] [4][3] [2][0]
vpandn @T[8],$A31,$A31 # tgting [3][1] [1][2] [4][3] [2][4]
vpandn @T[7],$A41,$A41 # tgting [3][2] [1][4] [4][1] [2][3]
vpblendd \$0b00001100,@T[2],@T[5],$A11 # [4][0] [2][3]
vpblendd \$0b00001100,@T[5],@T[3],@T[8] # [4][1] [2][4]
vpxor @T[3],$A31,$A31
vpblendd \$0b00110000,@T[3],$A11,$A11 # [1][2] [4][0] [2][3]
vpblendd \$0b00110000,@T[4],@T[8],@T[8] # [1][3] [4][1] [2][4]
vpxor @T[5],$A41,$A41
vpblendd \$0b11000000,@T[4],$A11,$A11 # [3][4] [1][2] [4][0] [2][3]
vpblendd \$0b11000000,@T[2],@T[8],@T[8] # [3][0] [1][3] [4][1] [2][4]
vpandn @T[8],$A11,$A11 # tgting [3][3] [1][1] [4][4] [2][2]
vpxor @T[6],$A11,$A11
vpermq \$0b00011110,@T[1],$A21 # [0][1] [0][2] [0][4] [0][3]
vpblendd \$0b00110000,$A00,$A21,@T[8] # [0][1] [0][0] [0][4] [0][3]
vpermq \$0b00111001,@T[1],$A01 # [0][1] [0][4] [0][3] [0][2]
vpblendd \$0b11000000,$A00,$A01,$A01 # [0][0] [0][4] [0][3] [0][2]
vpandn @T[8],$A01,$A01 # tgting [0][4] [0][3] [0][2] [0][1]
vpblendd \$0b00001100,@T[5],@T[4],$A20 # [4][1] [2][1]
vpblendd \$0b00001100,@T[4],@T[6],@T[7] # [4][2] [2][2]
vpblendd \$0b00110000,@T[6],$A20,$A20 # [1][1] [4][1] [2][1]
vpblendd \$0b00110000,@T[3],@T[7],@T[7] # [1][2] [4][2] [2][2]
vpblendd \$0b11000000,@T[3],$A20,$A20 # [3][1] [1][1] [4][1] [2][1]
vpblendd \$0b11000000,@T[5],@T[7],@T[7] # [3][2] [1][2] [4][2] [2][2]
vpandn @T[7],$A20,$A20 # tgting [3][0] [1][0] [4][0] [2][0]
vpxor @T[2],$A20,$A20
vpermq \$0b00000000,@T[0],@T[0] # [0][0] [0][0] [0][0] [0][0]
vpermq \$0b00011011,$A31,$A31 # post-Chi shuffle
vpermq \$0b10001101,$A41,$A41
vpermq \$0b01110010,$A11,$A11
vpblendd \$0b00001100,@T[3],@T[6],$A21 # [4][3] [2][2]
vpblendd \$0b00001100,@T[6],@T[5],@T[7] # [4][4] [2][3]
vpblendd \$0b00110000,@T[5],$A21,$A21 # [1][4] [4][3] [2][2]
vpblendd \$0b00110000,@T[2],@T[7],@T[7] # [1][0] [4][4] [2][3]
vpblendd \$0b11000000,@T[2],$A21,$A21 # [3][0] [1][4] [4][3] [2][2]
vpblendd \$0b11000000,@T[3],@T[7],@T[7] # [3][1] [1][0] [4][4] [2][3]
vpandn @T[7],$A21,$A21 # tgting [3][4] [1][3] [4][2] [2][1]
vpxor @T[0],$A00,$A00
vpxor @T[1],$A01,$A01
vpxor @T[4],$A21,$A21
######################################### Iota
vpxor (%r10),$A00,$A00
lea 32(%r10),%r10
dec %eax
jnz .Loop_avx2
ret
.size __KeccakF1600,.-__KeccakF1600
___
my ($A_flat,$inp,$len,$bsz) = ("%rdi","%rsi","%rdx","%rcx");
my $out = $inp; # in squeeze
$code.=<<___;
.globl SHA3_absorb
.type SHA3_absorb,\@function
.align 32
SHA3_absorb:
mov %rsp,%r11
lea -240(%rsp),%rsp
and \$-32,%rsp
lea 96($A_flat),$A_flat
lea 96($inp),$inp
lea 96(%rsp),%r10
vzeroupper
vpbroadcastq -96($A_flat),$A00 # load A[5][5]
vmovdqu 8+32*0-96($A_flat),$A01
vmovdqu 8+32*1-96($A_flat),$A20
vmovdqu 8+32*2-96($A_flat),$A31
vmovdqu 8+32*3-96($A_flat),$A21
vmovdqu 8+32*4-96($A_flat),$A41
vmovdqu 8+32*5-96($A_flat),$A11
vpxor @T[0],@T[0],@T[0]
vmovdqa @T[0],32*2-96(%r10) # zero transfer area on stack
vmovdqa @T[0],32*3-96(%r10)
vmovdqa @T[0],32*4-96(%r10)
vmovdqa @T[0],32*5-96(%r10)
vmovdqa @T[0],32*6-96(%r10)
.Loop_absorb_avx2:
mov $bsz,%rax
sub $bsz,$len
jc .Ldone_absorb_avx2
shr \$3,%eax
vpbroadcastq 0-96($inp),@T[0]
vmovdqu 8-96($inp),@T[1]
sub \$4,%eax
___
for(my $i=5; $i<25; $i++) {
$code.=<<___
dec %eax
jz .Labsorved_avx2
mov 8*$i-96($inp),%r8
mov %r8,$A_jagged[$i]-96(%r10)
___
}
$code.=<<___;
.Labsorved_avx2:
lea ($inp,$bsz),$inp
vpxor @T[0],$A00,$A00
vpxor @T[1],$A01,$A01
vpxor 32*2-96(%r10),$A20,$A20
vpxor 32*3-96(%r10),$A31,$A31
vpxor 32*4-96(%r10),$A21,$A21
vpxor 32*5-96(%r10),$A41,$A41
vpxor 32*6-96(%r10),$A11,$A11
call __KeccakF1600
lea 96(%rsp),%r10
jmp .Loop_absorb_avx2
.Ldone_absorb_avx2:
vmovq %xmm0,-96($A_flat)
vmovdqu $A01,8+32*0-96($A_flat)
vmovdqu $A20,8+32*1-96($A_flat)
vmovdqu $A31,8+32*2-96($A_flat)
vmovdqu $A21,8+32*3-96($A_flat)
vmovdqu $A41,8+32*4-96($A_flat)
vmovdqu $A11,8+32*5-96($A_flat)
vzeroupper
lea (%r11),%rsp
lea ($len,$bsz),%rax # return value
ret
.size SHA3_absorb,.-SHA3_absorb
.globl SHA3_squeeze
.type SHA3_squeeze,\@function
.align 32
SHA3_squeeze:
mov %rsp,%r11
lea 96($A_flat),$A_flat
shr \$3,$bsz
vzeroupper
vpbroadcastq -96($A_flat),$A00
vpxor @T[0],@T[0],@T[0]
vmovdqu 8+32*0-96($A_flat),$A01
vmovdqu 8+32*1-96($A_flat),$A20
vmovdqu 8+32*2-96($A_flat),$A31
vmovdqu 8+32*3-96($A_flat),$A21
vmovdqu 8+32*4-96($A_flat),$A41
vmovdqu 8+32*5-96($A_flat),$A11
mov $bsz,%rax
.Loop_squeeze_avx2:
mov @A_jagged[$i]-96($A_flat),%r8
___
for (my $i=0; $i<25; $i++) {
$code.=<<___;
sub \$8,$len
jc .Ltail_squeeze_avx2
mov %r8,($out)
lea 8($out),$out
je .Ldone_squeeze_avx2
dec %eax
je .Lextend_output_avx2
mov @A_jagged[$i+1]-120($A_flat),%r8
___
}
$code.=<<___;
.Lextend_output_avx2:
call __KeccakF1600
vmovq %xmm0,-96($A_flat)
vmovdqu $A01,8+32*0-96($A_flat)
vmovdqu $A20,8+32*1-96($A_flat)
vmovdqu $A31,8+32*2-96($A_flat)
vmovdqu $A21,8+32*3-96($A_flat)
vmovdqu $A41,8+32*4-96($A_flat)
vmovdqu $A11,8+32*5-96($A_flat)
mov $bsz,%rax
jmp .Loop_squeeze_avx2
.Ltail_squeeze_avx2:
add \$8,$len
.Loop_tail_avx2:
mov %r8b,($out)
lea 1($out),$out
shr \$8,%r8
dec $len
jnz .Loop_tail_avx2
.Ldone_squeeze_avx2:
vzeroupper
lea (%r11),%rsp
ret
.size SHA3_squeeze,.-SHA3_squeeze
.align 64
rhotates_left:
.quad 3, 18, 36, 41 # [2][0] [4][0] [1][0] [3][0]
.quad 1, 62, 28, 27 # [0][1] [0][2] [0][3] [0][4]
.quad 45, 6, 56, 39 # [3][1] [1][2] [4][3] [2][4]
.quad 10, 61, 55, 8 # [2][1] [4][2] [1][3] [3][4]
.quad 2, 15, 25, 20 # [4][1] [3][2] [2][3] [1][4]
.quad 44, 43, 21, 14 # [1][1] [2][2] [3][3] [4][4]
rhotates_right:
.quad 64-3, 64-18, 64-36, 64-41
.quad 64-1, 64-62, 64-28, 64-27
.quad 64-45, 64-6, 64-56, 64-39
.quad 64-10, 64-61, 64-55, 64-8
.quad 64-2, 64-15, 64-25, 64-20
.quad 64-44, 64-43, 64-21, 64-14
iotas:
.quad 0x0000000000000001, 0x0000000000000001, 0x0000000000000001, 0x0000000000000001
.quad 0x0000000000008082, 0x0000000000008082, 0x0000000000008082, 0x0000000000008082
.quad 0x800000000000808a, 0x800000000000808a, 0x800000000000808a, 0x800000000000808a
.quad 0x8000000080008000, 0x8000000080008000, 0x8000000080008000, 0x8000000080008000
.quad 0x000000000000808b, 0x000000000000808b, 0x000000000000808b, 0x000000000000808b
.quad 0x0000000080000001, 0x0000000080000001, 0x0000000080000001, 0x0000000080000001
.quad 0x8000000080008081, 0x8000000080008081, 0x8000000080008081, 0x8000000080008081
.quad 0x8000000000008009, 0x8000000000008009, 0x8000000000008009, 0x8000000000008009
.quad 0x000000000000008a, 0x000000000000008a, 0x000000000000008a, 0x000000000000008a
.quad 0x0000000000000088, 0x0000000000000088, 0x0000000000000088, 0x0000000000000088
.quad 0x0000000080008009, 0x0000000080008009, 0x0000000080008009, 0x0000000080008009
.quad 0x000000008000000a, 0x000000008000000a, 0x000000008000000a, 0x000000008000000a
.quad 0x000000008000808b, 0x000000008000808b, 0x000000008000808b, 0x000000008000808b
.quad 0x800000000000008b, 0x800000000000008b, 0x800000000000008b, 0x800000000000008b
.quad 0x8000000000008089, 0x8000000000008089, 0x8000000000008089, 0x8000000000008089
.quad 0x8000000000008003, 0x8000000000008003, 0x8000000000008003, 0x8000000000008003
.quad 0x8000000000008002, 0x8000000000008002, 0x8000000000008002, 0x8000000000008002
.quad 0x8000000000000080, 0x8000000000000080, 0x8000000000000080, 0x8000000000000080
.quad 0x000000000000800a, 0x000000000000800a, 0x000000000000800a, 0x000000000000800a
.quad 0x800000008000000a, 0x800000008000000a, 0x800000008000000a, 0x800000008000000a
.quad 0x8000000080008081, 0x8000000080008081, 0x8000000080008081, 0x8000000080008081
.quad 0x8000000000008080, 0x8000000000008080, 0x8000000000008080, 0x8000000000008080
.quad 0x0000000080000001, 0x0000000080000001, 0x0000000080000001, 0x0000000080000001
.quad 0x8000000080008008, 0x8000000080008008, 0x8000000080008008, 0x8000000080008008
.asciz "Keccak-1600 absorb and squeeze for AVX2, CRYPTOGAMS by <appro\@openssl.org>"
___
print $code;
close STDOUT;
Reference in a new issue View git blame Copy permalink