|
Packit |
6c4009 |
/* Function cosf vectorized with AVX2.
|
|
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 |
|
|
Packit |
6c4009 |
#include <sysdep.h>
|
|
Packit |
6c4009 |
#include "svml_s_trig_data.h"
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (_ZGVdN8v_cosf_avx2)
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
ALGORITHM DESCRIPTION:
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
1) Range reduction to [-Pi/2; +Pi/2] interval
|
|
Packit |
6c4009 |
a) We remove sign using AND operation
|
|
Packit |
6c4009 |
b) Add Pi/2 value to argument X for Cos to Sin transformation
|
|
Packit |
6c4009 |
c) Getting octant Y by 1/Pi multiplication
|
|
Packit |
6c4009 |
d) Add "Right Shifter" value
|
|
Packit |
6c4009 |
e) Treat obtained value as integer for destination sign setting.
|
|
Packit |
6c4009 |
Shift first bit of this value to the last (sign) position
|
|
Packit |
6c4009 |
f) Subtract "Right Shifter" value
|
|
Packit |
6c4009 |
g) Subtract 0.5 from result for octant correction
|
|
Packit |
6c4009 |
h) Subtract Y*PI from X argument, where PI divided to 4 parts:
|
|
Packit |
6c4009 |
X = X - Y*PI1 - Y*PI2 - Y*PI3 - Y*PI4;
|
|
Packit |
6c4009 |
2) Polynomial (minimax for sin within [-Pi/2; +Pi/2] interval)
|
|
Packit |
6c4009 |
a) Calculate X^2 = X * X
|
|
Packit |
6c4009 |
b) Calculate polynomial:
|
|
Packit |
6c4009 |
R = X + X * X^2 * (A3 + x^2 * (A5 + .....
|
|
Packit |
6c4009 |
3) Destination sign setting
|
|
Packit |
6c4009 |
a) Set shifted destination sign using XOR operation:
|
|
Packit |
6c4009 |
R = XOR( R, S );
|
|
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 $448, %rsp
|
|
Packit |
6c4009 |
movq __svml_s_trig_data@GOTPCREL(%rip), %rax
|
|
Packit |
6c4009 |
vmovaps %ymm0, %ymm2
|
|
Packit |
6c4009 |
vmovups __sRShifter(%rax), %ymm5
|
|
Packit |
6c4009 |
vmovups __sPI1_FMA(%rax), %ymm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* b) Add Pi/2 value to argument X for Cos to Sin transformation */
|
|
Packit |
6c4009 |
vaddps __sHalfPI(%rax), %ymm2, %ymm4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
1) Range reduction to [-Pi/2; +Pi/2] interval
|
|
Packit |
6c4009 |
c) Getting octant Y by 1/Pi multiplication
|
|
Packit |
6c4009 |
d) Add "Right Shifter" (0x4B000000) value
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
vfmadd132ps __sInvPI(%rax), %ymm5, %ymm4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* f) Subtract "Right Shifter" (0x4B000000) value */
|
|
Packit |
6c4009 |
vsubps %ymm5, %ymm4, %ymm6
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
e) Treat obtained value as integer for destination sign setting.
|
|
Packit |
6c4009 |
Shift first bit of this value to the last (sign) position (S << 31)
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
vpslld $31, %ymm4, %ymm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* g) Subtract 0.5 from result for octant correction */
|
|
Packit |
6c4009 |
vsubps __sOneHalf(%rax), %ymm6, %ymm4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Check for large and special arguments */
|
|
Packit |
6c4009 |
vandps __sAbsMask(%rax), %ymm2, %ymm3
|
|
Packit |
6c4009 |
vcmpnle_uqps __sRangeReductionVal(%rax), %ymm3, %ymm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
h) Subtract Y*PI from X argument, where PI divided to 4 parts:
|
|
Packit |
6c4009 |
X = X - Y*PI1 - Y*PI2 - Y*PI3
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
vmovaps %ymm2, %ymm3
|
|
Packit |
6c4009 |
vfnmadd231ps %ymm4, %ymm7, %ymm3
|
|
Packit |
6c4009 |
vfnmadd231ps __sPI2_FMA(%rax), %ymm4, %ymm3
|
|
Packit |
6c4009 |
vfnmadd132ps __sPI3_FMA(%rax), %ymm3, %ymm4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* a) Calculate X^2 = X * X */
|
|
Packit |
6c4009 |
vmulps %ymm4, %ymm4, %ymm5
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
3) Destination sign setting
|
|
Packit |
6c4009 |
a) Set shifted destination sign using XOR operation:
|
|
Packit |
6c4009 |
R = XOR( R, S );
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
vxorps %ymm0, %ymm4, %ymm6
|
|
Packit |
6c4009 |
vmovups __sA9_FMA(%rax), %ymm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
b) Calculate polynomial:
|
|
Packit |
6c4009 |
R = X + X * X^2 * (A3 + x^2 * (A5 + x^2 * (A7 + x^2 * (A9))))
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
vfmadd213ps __sA7_FMA(%rax), %ymm5, %ymm0
|
|
Packit |
6c4009 |
vfmadd213ps __sA5_FMA(%rax), %ymm5, %ymm0
|
|
Packit |
6c4009 |
vfmadd213ps __sA3(%rax), %ymm5, %ymm0
|
|
Packit |
6c4009 |
vmulps %ymm5, %ymm0, %ymm0
|
|
Packit |
6c4009 |
vmovmskps %ymm1, %ecx
|
|
Packit |
6c4009 |
vfmadd213ps %ymm6, %ymm6, %ymm0
|
|
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 |
vmovups %ymm2, 320(%rsp)
|
|
Packit |
6c4009 |
vmovups %ymm0, 384(%rsp)
|
|
Packit |
6c4009 |
je .LBL_1_2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
xorb %dl, %dl
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
vmovups %ymm8, 224(%rsp)
|
|
Packit |
6c4009 |
vmovups %ymm9, 192(%rsp)
|
|
Packit |
6c4009 |
vmovups %ymm10, 160(%rsp)
|
|
Packit |
6c4009 |
vmovups %ymm11, 128(%rsp)
|
|
Packit |
6c4009 |
vmovups %ymm12, 96(%rsp)
|
|
Packit |
6c4009 |
vmovups %ymm13, 64(%rsp)
|
|
Packit |
6c4009 |
vmovups %ymm14, 32(%rsp)
|
|
Packit |
6c4009 |
vmovups %ymm15, (%rsp)
|
|
Packit |
6c4009 |
movq %rsi, 264(%rsp)
|
|
Packit |
6c4009 |
movq %rdi, 256(%rsp)
|
|
Packit |
6c4009 |
movq %r12, 296(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (12, 296)
|
|
Packit |
6c4009 |
movb %dl, %r12b
|
|
Packit |
6c4009 |
movq %r13, 288(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (13, 288)
|
|
Packit |
6c4009 |
movl %ecx, %r13d
|
|
Packit |
6c4009 |
movq %r14, 280(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (14, 280)
|
|
Packit |
6c4009 |
movl %eax, %r14d
|
|
Packit |
6c4009 |
movq %r15, 272(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (15, 272)
|
|
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 |
vmovups 224(%rsp), %ymm8
|
|
Packit |
6c4009 |
vmovups 192(%rsp), %ymm9
|
|
Packit |
6c4009 |
vmovups 160(%rsp), %ymm10
|
|
Packit |
6c4009 |
vmovups 128(%rsp), %ymm11
|
|
Packit |
6c4009 |
vmovups 96(%rsp), %ymm12
|
|
Packit |
6c4009 |
vmovups 64(%rsp), %ymm13
|
|
Packit |
6c4009 |
vmovups 32(%rsp), %ymm14
|
|
Packit |
6c4009 |
vmovups (%rsp), %ymm15
|
|
Packit |
6c4009 |
vmovups 384(%rsp), %ymm0
|
|
Packit |
6c4009 |
movq 264(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 256(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 296(%rsp), %r12
|
|
Packit |
6c4009 |
cfi_restore (%r12)
|
|
Packit |
6c4009 |
movq 288(%rsp), %r13
|
|
Packit |
6c4009 |
cfi_restore (%r13)
|
|
Packit |
6c4009 |
movq 280(%rsp), %r14
|
|
Packit |
6c4009 |
cfi_restore (%r14)
|
|
Packit |
6c4009 |
movq 272(%rsp), %r15
|
|
Packit |
6c4009 |
cfi_restore (%r15)
|
|
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 |
vmovss 324(%rsp,%r15,8), %xmm0
|
|
Packit |
6c4009 |
vzeroupper
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(cosf)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
vmovss %xmm0, 388(%rsp,%r15,8)
|
|
Packit |
6c4009 |
jmp .LBL_1_8
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_12:
|
|
Packit |
6c4009 |
movzbl %r12b, %r15d
|
|
Packit |
6c4009 |
vmovss 320(%rsp,%r15,8), %xmm0
|
|
Packit |
6c4009 |
vzeroupper
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(cosf)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
vmovss %xmm0, 384(%rsp,%r15,8)
|
|
Packit |
6c4009 |
jmp .LBL_1_7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
END (_ZGVdN8v_cosf_avx2)
|