|
Packit |
5c3484 |
dnl AMD64 mpn_popcount, mpn_hamdist -- population count and hamming distance.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2004, 2005, 2007, 2010-2012 Free Software Foundation, Inc.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl This file is part of the GNU MP Library.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl The GNU MP Library is free software; you can redistribute it and/or modify
|
|
Packit |
5c3484 |
dnl it under the terms of either:
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl * the GNU Lesser General Public License as published by the Free
|
|
Packit |
5c3484 |
dnl Software Foundation; either version 3 of the License, or (at your
|
|
Packit |
5c3484 |
dnl option) any later version.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl or
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl * the GNU General Public License as published by the Free Software
|
|
Packit |
5c3484 |
dnl Foundation; either version 2 of the License, or (at your option) any
|
|
Packit |
5c3484 |
dnl later version.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl or both in parallel, as here.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl The GNU MP Library is distributed in the hope that it will be useful, but
|
|
Packit |
5c3484 |
dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
|
|
Packit |
5c3484 |
dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
Packit |
5c3484 |
dnl for more details.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl You should have received copies of the GNU General Public License and the
|
|
Packit |
5c3484 |
dnl GNU Lesser General Public License along with the GNU MP Library. If not,
|
|
Packit |
5c3484 |
dnl see https://www.gnu.org/licenses/.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
include(`../config.m4')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C popcount hamdist
|
|
Packit |
5c3484 |
C cycles/limb cycles/limb
|
|
Packit |
5c3484 |
C AMD K8,K9 6 7
|
|
Packit |
5c3484 |
C AMD K10 6 7
|
|
Packit |
5c3484 |
C Intel P4 12 14.3
|
|
Packit |
5c3484 |
C Intel core2 7 8
|
|
Packit |
5c3484 |
C Intel corei ? 7.3
|
|
Packit |
5c3484 |
C Intel atom 16.5 17.5
|
|
Packit |
5c3484 |
C VIA nano 8.75 10.4
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C TODO
|
|
Packit |
5c3484 |
C * Tune. It should be possible to reach 5 c/l for popcount and 6 c/l for
|
|
Packit |
5c3484 |
C hamdist for K8/K9.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`OPERATION_popcount',`
|
|
Packit |
5c3484 |
define(`func',`mpn_popcount')
|
|
Packit |
5c3484 |
define(`up', `%rdi')
|
|
Packit |
5c3484 |
define(`n', `%rsi')
|
|
Packit |
5c3484 |
define(`h55555555', `%r10')
|
|
Packit |
5c3484 |
define(`h33333333', `%r11')
|
|
Packit |
5c3484 |
define(`h0f0f0f0f', `%rcx')
|
|
Packit |
5c3484 |
define(`h01010101', `%rdx')
|
|
Packit |
5c3484 |
define(`POP', `$1')
|
|
Packit |
5c3484 |
define(`HAM', `dnl')
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
ifdef(`OPERATION_hamdist',`
|
|
Packit |
5c3484 |
define(`func',`mpn_hamdist')
|
|
Packit |
5c3484 |
define(`up', `%rdi')
|
|
Packit |
5c3484 |
define(`vp', `%rsi')
|
|
Packit |
5c3484 |
define(`n', `%rdx')
|
|
Packit |
5c3484 |
define(`h55555555', `%r10')
|
|
Packit |
5c3484 |
define(`h33333333', `%r11')
|
|
Packit |
5c3484 |
define(`h0f0f0f0f', `%rcx')
|
|
Packit |
5c3484 |
define(`h01010101', `%r14')
|
|
Packit |
5c3484 |
define(`POP', `dnl')
|
|
Packit |
5c3484 |
define(`HAM', `$1')
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
MULFUNC_PROLOGUE(mpn_popcount mpn_hamdist)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ABI_SUPPORT(DOS64)
|
|
Packit |
5c3484 |
ABI_SUPPORT(STD64)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
PROLOGUE(func)
|
|
Packit |
5c3484 |
POP(` FUNC_ENTRY(2) ')
|
|
Packit |
5c3484 |
HAM(` FUNC_ENTRY(3) ')
|
|
Packit |
5c3484 |
push %r12
|
|
Packit |
5c3484 |
push %r13
|
|
Packit |
5c3484 |
HAM(` push %r14 ')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov $0x5555555555555555, h55555555
|
|
Packit |
5c3484 |
mov $0x3333333333333333, h33333333
|
|
Packit |
5c3484 |
mov $0x0f0f0f0f0f0f0f0f, h0f0f0f0f
|
|
Packit |
5c3484 |
mov $0x0101010101010101, h01010101
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
lea (up,n,8), up
|
|
Packit |
5c3484 |
HAM(` lea (vp,n,8), vp ')
|
|
Packit |
5c3484 |
neg n
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
xor R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
bt $0, R32(n)
|
|
Packit |
5c3484 |
jnc L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up,n,8), %r8
|
|
Packit |
5c3484 |
HAM(` xor (vp,n,8), %r8 ')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov %r8, %r9
|
|
Packit |
5c3484 |
shr %r8
|
|
Packit |
5c3484 |
and h55555555, %r8
|
|
Packit |
5c3484 |
sub %r8, %r9
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov %r9, %r8
|
|
Packit |
5c3484 |
shr $2, %r9
|
|
Packit |
5c3484 |
and h33333333, %r8
|
|
Packit |
5c3484 |
and h33333333, %r9
|
|
Packit |
5c3484 |
add %r8, %r9 C 16 4-bit fields (0..4)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov %r9, %r8
|
|
Packit |
5c3484 |
shr $4, %r9
|
|
Packit |
5c3484 |
and h0f0f0f0f, %r8
|
|
Packit |
5c3484 |
and h0f0f0f0f, %r9
|
|
Packit |
5c3484 |
add %r8, %r9 C 8 8-bit fields (0..16)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
imul h01010101, %r9 C sum the 8 fields in high 8 bits
|
|
Packit |
5c3484 |
shr $56, %r9
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov %r9, %rax C add to total
|
|
Packit |
5c3484 |
add $1, n
|
|
Packit |
5c3484 |
jz L(end)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(top): mov (up,n,8), %r8
|
|
Packit |
5c3484 |
mov 8(up,n,8), %r12
|
|
Packit |
5c3484 |
HAM(` xor (vp,n,8), %r8 ')
|
|
Packit |
5c3484 |
HAM(` xor 8(vp,n,8), %r12 ')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov %r8, %r9
|
|
Packit |
5c3484 |
mov %r12, %r13
|
|
Packit |
5c3484 |
shr %r8
|
|
Packit |
5c3484 |
shr %r12
|
|
Packit |
5c3484 |
and h55555555, %r8
|
|
Packit |
5c3484 |
and h55555555, %r12
|
|
Packit |
5c3484 |
sub %r8, %r9
|
|
Packit |
5c3484 |
sub %r12, %r13
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov %r9, %r8
|
|
Packit |
5c3484 |
mov %r13, %r12
|
|
Packit |
5c3484 |
shr $2, %r9
|
|
Packit |
5c3484 |
shr $2, %r13
|
|
Packit |
5c3484 |
and h33333333, %r8
|
|
Packit |
5c3484 |
and h33333333, %r9
|
|
Packit |
5c3484 |
and h33333333, %r12
|
|
Packit |
5c3484 |
and h33333333, %r13
|
|
Packit |
5c3484 |
add %r8, %r9 C 16 4-bit fields (0..4)
|
|
Packit |
5c3484 |
add %r12, %r13 C 16 4-bit fields (0..4)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
add %r13, %r9 C 16 4-bit fields (0..8)
|
|
Packit |
5c3484 |
mov %r9, %r8
|
|
Packit |
5c3484 |
shr $4, %r9
|
|
Packit |
5c3484 |
and h0f0f0f0f, %r8
|
|
Packit |
5c3484 |
and h0f0f0f0f, %r9
|
|
Packit |
5c3484 |
add %r8, %r9 C 8 8-bit fields (0..16)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
imul h01010101, %r9 C sum the 8 fields in high 8 bits
|
|
Packit |
5c3484 |
shr $56, %r9
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
add %r9, %rax C add to total
|
|
Packit |
5c3484 |
add $2, n
|
|
Packit |
5c3484 |
jnc L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end):
|
|
Packit |
5c3484 |
HAM(` pop %r14 ')
|
|
Packit |
5c3484 |
pop %r13
|
|
Packit |
5c3484 |
pop %r12
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|