|
Packit |
6c4009 |
/* Function expf vectorized with SSE4.
|
|
Packit |
6c4009 |
Copyright (C) 2014-2018 Free Software Foundation, Inc.
|
|
Packit |
6c4009 |
This file is part of the GNU C Library.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is free software; you can redistribute it and/or
|
|
Packit |
6c4009 |
modify it under the terms of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License as published by the Free Software Foundation; either
|
|
Packit |
6c4009 |
version 2.1 of the License, or (at your option) any later version.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is distributed in the hope that it will be useful,
|
|
Packit |
6c4009 |
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
Packit |
6c4009 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
Packit |
6c4009 |
Lesser General Public License for more details.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
You should have received a copy of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License along with the GNU C Library; if not, see
|
|
Packit |
6c4009 |
<http://www.gnu.org/licenses/>. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#include <sysdep.h>
|
|
Packit |
6c4009 |
#include "svml_s_expf_data.h"
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (_ZGVbN4v_expf_sse4)
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
ALGORITHM DESCRIPTION:
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
Argument representation:
|
|
Packit |
6c4009 |
M = rint(X*2^k/ln2) = 2^k*N+j
|
|
Packit |
6c4009 |
X = M*ln2/2^k + r = N*ln2 + ln2*(j/2^k) + r
|
|
Packit |
6c4009 |
then -ln2/2^(k+1) < r < ln2/2^(k+1)
|
|
Packit |
6c4009 |
Alternatively:
|
|
Packit |
6c4009 |
M = trunc(X*2^k/ln2)
|
|
Packit |
6c4009 |
then 0 < r < ln2/2^k
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
Result calculation:
|
|
Packit |
6c4009 |
exp(X) = exp(N*ln2 + ln2*(j/2^k) + r)
|
|
Packit |
6c4009 |
= 2^N * 2^(j/2^k) * exp(r)
|
|
Packit |
6c4009 |
2^N is calculated by bit manipulation
|
|
Packit |
6c4009 |
2^(j/2^k) is computed from table lookup
|
|
Packit |
6c4009 |
exp(r) is approximated by polynomial
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The table lookup is skipped if k = 0.
|
|
Packit |
6c4009 |
For low accuracy approximation, exp(r) ~ 1 or 1+r. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pushq %rbp
|
|
Packit |
6c4009 |
cfi_adjust_cfa_offset (8)
|
|
Packit |
6c4009 |
cfi_rel_offset (%rbp, 0)
|
|
Packit |
6c4009 |
movq %rsp, %rbp
|
|
Packit |
6c4009 |
cfi_def_cfa_register (%rbp)
|
|
Packit |
6c4009 |
andq $-64, %rsp
|
|
Packit |
6c4009 |
subq $320, %rsp
|
|
Packit |
6c4009 |
movaps %xmm0, %xmm5
|
|
Packit |
6c4009 |
movq __svml_sexp_data@GOTPCREL(%rip), %rax
|
|
Packit |
6c4009 |
movups __sInvLn2(%rax), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* m = x*2^k/ln2 + shifter */
|
|
Packit |
6c4009 |
mulps %xmm5, %xmm0
|
|
Packit |
6c4009 |
movups __sShifter(%rax), %xmm6
|
|
Packit |
6c4009 |
movups __sLn2hi(%rax), %xmm4
|
|
Packit |
6c4009 |
addps %xmm6, %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* n = m - shifter = rint(x*2^k/ln2) */
|
|
Packit |
6c4009 |
movaps %xmm0, %xmm2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* remove sign of x by "and" operation */
|
|
Packit |
6c4009 |
movdqu __iAbsMask(%rax), %xmm7
|
|
Packit |
6c4009 |
subps %xmm6, %xmm2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* r = x-n*ln2_hi/2^k */
|
|
Packit |
6c4009 |
mulps %xmm2, %xmm4
|
|
Packit |
6c4009 |
pand %xmm5, %xmm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* compare against threshold */
|
|
Packit |
6c4009 |
pcmpgtd __iDomainRange(%rax), %xmm7
|
|
Packit |
6c4009 |
movups __sLn2lo(%rax), %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* set mask for overflow/underflow */
|
|
Packit |
6c4009 |
movmskps %xmm7, %ecx
|
|
Packit |
6c4009 |
movaps %xmm5, %xmm7
|
|
Packit |
6c4009 |
movups __sPC5(%rax), %xmm3
|
|
Packit |
6c4009 |
subps %xmm4, %xmm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* r = r-n*ln2_lo/2^k = x - n*ln2/2^k */
|
|
Packit |
6c4009 |
mulps %xmm1, %xmm2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* compute 2^N with "shift" */
|
|
Packit |
6c4009 |
movdqu __iBias(%rax), %xmm6
|
|
Packit |
6c4009 |
subps %xmm2, %xmm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* c5*r+c4 */
|
|
Packit |
6c4009 |
mulps %xmm7, %xmm3
|
|
Packit |
6c4009 |
paddd %xmm6, %xmm0
|
|
Packit |
6c4009 |
pslld $23, %xmm0
|
|
Packit |
6c4009 |
addps __sPC4(%rax), %xmm3
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* (c5*r+c4)*r+c3 */
|
|
Packit |
6c4009 |
mulps %xmm7, %xmm3
|
|
Packit |
6c4009 |
addps __sPC3(%rax), %xmm3
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* ((c5*r+c4)*r+c3)*r+c2 */
|
|
Packit |
6c4009 |
mulps %xmm7, %xmm3
|
|
Packit |
6c4009 |
addps __sPC2(%rax), %xmm3
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* (((c5*r+c4)*r+c3)*r+c2)*r+c1 */
|
|
Packit |
6c4009 |
mulps %xmm7, %xmm3
|
|
Packit |
6c4009 |
addps __sPC1(%rax), %xmm3
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* exp(r) = ((((c5*r+c4)*r+c3)*r+c2)*r+c1)*r+c0 */
|
|
Packit |
6c4009 |
mulps %xmm3, %xmm7
|
|
Packit |
6c4009 |
addps __sPC0(%rax), %xmm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* 2^N*exp(r) */
|
|
Packit |
6c4009 |
mulps %xmm7, %xmm0
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
jne .LBL_1_3
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_2:
|
|
Packit |
6c4009 |
cfi_remember_state
|
|
Packit |
6c4009 |
movq %rbp, %rsp
|
|
Packit |
6c4009 |
cfi_def_cfa_register (%rsp)
|
|
Packit |
6c4009 |
popq %rbp
|
|
Packit |
6c4009 |
cfi_adjust_cfa_offset (-8)
|
|
Packit |
6c4009 |
cfi_restore (%rbp)
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_3:
|
|
Packit |
6c4009 |
cfi_restore_state
|
|
Packit |
6c4009 |
movups %xmm5, 192(%rsp)
|
|
Packit |
6c4009 |
movups %xmm0, 256(%rsp)
|
|
Packit |
6c4009 |
je .LBL_1_2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
xorb %dl, %dl
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movups %xmm8, 112(%rsp)
|
|
Packit |
6c4009 |
movups %xmm9, 96(%rsp)
|
|
Packit |
6c4009 |
movups %xmm10, 80(%rsp)
|
|
Packit |
6c4009 |
movups %xmm11, 64(%rsp)
|
|
Packit |
6c4009 |
movups %xmm12, 48(%rsp)
|
|
Packit |
6c4009 |
movups %xmm13, 32(%rsp)
|
|
Packit |
6c4009 |
movups %xmm14, 16(%rsp)
|
|
Packit |
6c4009 |
movups %xmm15, (%rsp)
|
|
Packit |
6c4009 |
movq %rsi, 136(%rsp)
|
|
Packit |
6c4009 |
movq %rdi, 128(%rsp)
|
|
Packit |
6c4009 |
movq %r12, 168(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (12, 168)
|
|
Packit |
6c4009 |
movb %dl, %r12b
|
|
Packit |
6c4009 |
movq %r13, 160(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (13, 160)
|
|
Packit |
6c4009 |
movl %ecx, %r13d
|
|
Packit |
6c4009 |
movq %r14, 152(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (14, 152)
|
|
Packit |
6c4009 |
movl %eax, %r14d
|
|
Packit |
6c4009 |
movq %r15, 144(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (15, 144)
|
|
Packit |
6c4009 |
cfi_remember_state
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_6:
|
|
Packit |
6c4009 |
btl %r14d, %r13d
|
|
Packit |
6c4009 |
jc .LBL_1_12
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_7:
|
|
Packit |
6c4009 |
lea 1(%r14), %esi
|
|
Packit |
6c4009 |
btl %esi, %r13d
|
|
Packit |
6c4009 |
jc .LBL_1_10
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_8:
|
|
Packit |
6c4009 |
incb %r12b
|
|
Packit |
6c4009 |
addl $2, %r14d
|
|
Packit |
6c4009 |
cmpb $16, %r12b
|
|
Packit |
6c4009 |
jb .LBL_1_6
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movups 112(%rsp), %xmm8
|
|
Packit |
6c4009 |
movups 96(%rsp), %xmm9
|
|
Packit |
6c4009 |
movups 80(%rsp), %xmm10
|
|
Packit |
6c4009 |
movups 64(%rsp), %xmm11
|
|
Packit |
6c4009 |
movups 48(%rsp), %xmm12
|
|
Packit |
6c4009 |
movups 32(%rsp), %xmm13
|
|
Packit |
6c4009 |
movups 16(%rsp), %xmm14
|
|
Packit |
6c4009 |
movups (%rsp), %xmm15
|
|
Packit |
6c4009 |
movq 136(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 128(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 168(%rsp), %r12
|
|
Packit |
6c4009 |
cfi_restore (%r12)
|
|
Packit |
6c4009 |
movq 160(%rsp), %r13
|
|
Packit |
6c4009 |
cfi_restore (%r13)
|
|
Packit |
6c4009 |
movq 152(%rsp), %r14
|
|
Packit |
6c4009 |
cfi_restore (%r14)
|
|
Packit |
6c4009 |
movq 144(%rsp), %r15
|
|
Packit |
6c4009 |
cfi_restore (%r15)
|
|
Packit |
6c4009 |
movups 256(%rsp), %xmm0
|
|
Packit |
6c4009 |
jmp .LBL_1_2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_10:
|
|
Packit |
6c4009 |
cfi_restore_state
|
|
Packit |
6c4009 |
movzbl %r12b, %r15d
|
|
Packit |
6c4009 |
movss 196(%rsp,%r15,8), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(__expf_finite)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movss %xmm0, 260(%rsp,%r15,8)
|
|
Packit |
6c4009 |
jmp .LBL_1_8
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_12:
|
|
Packit |
6c4009 |
movzbl %r12b, %r15d
|
|
Packit |
6c4009 |
movss 192(%rsp,%r15,8), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(__expf_finite)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movss %xmm0, 256(%rsp,%r15,8)
|
|
Packit |
6c4009 |
jmp .LBL_1_7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
END (_ZGVbN4v_expf_sse4)
|