|
Packit |
6c4009 |
/* Function sincos vectorized with AVX-512. Wrapper to AVX2 version.
|
|
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_d_wrapper_impl.h"
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (_ZGVeN8vl8l8_sincos)
|
|
Packit |
6c4009 |
WRAPPER_IMPL_AVX512_fFF _ZGVdN4vl8l8_sincos
|
|
Packit |
6c4009 |
END (_ZGVeN8vl8l8_sincos)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* AVX512 ISA version as wrapper to AVX2 ISA version (for vector
|
|
Packit |
6c4009 |
function declared with #pragma omp declare simd notinbranch). */
|
|
Packit |
6c4009 |
.macro WRAPPER_IMPL_AVX512_fFF_vvv callee
|
|
Packit |
6c4009 |
#ifndef __ILP32__
|
|
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 |
vmovups %zmm0, 256(%rsp)
|
|
Packit |
6c4009 |
lea (%rsp), %rdi
|
|
Packit |
6c4009 |
vmovups %zmm1, 128(%rdi)
|
|
Packit |
6c4009 |
vmovups %zmm2, 192(%rdi)
|
|
Packit |
6c4009 |
lea 64(%rsp), %rsi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
vmovdqu 288(%rsp), %ymm0
|
|
Packit |
6c4009 |
lea 32(%rsp), %rdi
|
|
Packit |
6c4009 |
lea 96(%rsp), %rsi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
movq 128(%rsp), %rdx
|
|
Packit |
6c4009 |
movq 192(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 136(%rsp), %r8
|
|
Packit |
6c4009 |
movq 200(%rsp), %r10
|
|
Packit |
6c4009 |
movq (%rsp), %rax
|
|
Packit |
6c4009 |
movq 64(%rsp), %rcx
|
|
Packit |
6c4009 |
movq 8(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 72(%rsp), %r9
|
|
Packit |
6c4009 |
movq %rax, (%rdx)
|
|
Packit |
6c4009 |
movq %rcx, (%rsi)
|
|
Packit |
6c4009 |
movq 144(%rsp), %rax
|
|
Packit |
6c4009 |
movq 208(%rsp), %rcx
|
|
Packit |
6c4009 |
movq %rdi, (%r8)
|
|
Packit |
6c4009 |
movq %r9, (%r10)
|
|
Packit |
6c4009 |
movq 152(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 216(%rsp), %r9
|
|
Packit |
6c4009 |
movq 16(%rsp), %r11
|
|
Packit |
6c4009 |
movq 80(%rsp), %rdx
|
|
Packit |
6c4009 |
movq 24(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 88(%rsp), %r8
|
|
Packit |
6c4009 |
movq %r11, (%rax)
|
|
Packit |
6c4009 |
movq %rdx, (%rcx)
|
|
Packit |
6c4009 |
movq 160(%rsp), %r11
|
|
Packit |
6c4009 |
movq 224(%rsp), %rdx
|
|
Packit |
6c4009 |
movq %rsi, (%rdi)
|
|
Packit |
6c4009 |
movq %r8, (%r9)
|
|
Packit |
6c4009 |
movq 168(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 232(%rsp), %r8
|
|
Packit |
6c4009 |
movq 32(%rsp), %r10
|
|
Packit |
6c4009 |
movq 96(%rsp), %rax
|
|
Packit |
6c4009 |
movq 40(%rsp), %rcx
|
|
Packit |
6c4009 |
movq 104(%rsp), %rdi
|
|
Packit |
6c4009 |
movq %r10, (%r11)
|
|
Packit |
6c4009 |
movq %rax, (%rdx)
|
|
Packit |
6c4009 |
movq 176(%rsp), %r10
|
|
Packit |
6c4009 |
movq 240(%rsp), %rax
|
|
Packit |
6c4009 |
movq %rcx, (%rsi)
|
|
Packit |
6c4009 |
movq %rdi, (%r8)
|
|
Packit |
6c4009 |
movq 184(%rsp), %rcx
|
|
Packit |
6c4009 |
movq 248(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 48(%rsp), %r9
|
|
Packit |
6c4009 |
movq 112(%rsp), %r11
|
|
Packit |
6c4009 |
movq 56(%rsp), %rdx
|
|
Packit |
6c4009 |
movq 120(%rsp), %rsi
|
|
Packit |
6c4009 |
movq %r9, (%r10)
|
|
Packit |
6c4009 |
movq %r11, (%rax)
|
|
Packit |
6c4009 |
movq %rdx, (%rcx)
|
|
Packit |
6c4009 |
movq %rsi, (%rdi)
|
|
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 |
#else
|
|
Packit |
6c4009 |
leal 8(%rsp), %r10d
|
|
Packit |
6c4009 |
.cfi_def_cfa 10, 0
|
|
Packit |
6c4009 |
andl $-64, %esp
|
|
Packit |
6c4009 |
pushq -8(%r10d)
|
|
Packit |
6c4009 |
pushq %rbp
|
|
Packit |
6c4009 |
.cfi_escape 0x10,0x6,0x2,0x76,0
|
|
Packit |
6c4009 |
movl %esp, %ebp
|
|
Packit |
6c4009 |
pushq %r12
|
|
Packit |
6c4009 |
leal -112(%rbp), %esi
|
|
Packit |
6c4009 |
pushq %r10
|
|
Packit |
6c4009 |
.cfi_escape 0xf,0x3,0x76,0x70,0x6
|
|
Packit |
6c4009 |
.cfi_escape 0x10,0xc,0x2,0x76,0x78
|
|
Packit |
6c4009 |
leal -176(%rbp), %edi
|
|
Packit |
6c4009 |
movq %rsi, %r12
|
|
Packit |
6c4009 |
pushq %rbx
|
|
Packit |
6c4009 |
.cfi_escape 0x10,0x3,0x2,0x76,0x68
|
|
Packit |
6c4009 |
movq %rdi, %rbx
|
|
Packit |
6c4009 |
subl $280, %esp
|
|
Packit |
6c4009 |
vmovdqa %ymm1, -208(%ebp)
|
|
Packit |
6c4009 |
vmovdqa %ymm2, -240(%ebp)
|
|
Packit |
6c4009 |
vmovapd %zmm0, -304(%ebp)
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
leal 32(%r12), %esi
|
|
Packit |
6c4009 |
vmovupd -272(%ebp), %ymm0
|
|
Packit |
6c4009 |
leal 32(%rbx), %edi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
movl -208(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -176(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -204(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -168(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -200(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -160(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -196(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -152(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -192(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -144(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -188(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -136(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -184(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -128(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -180(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -120(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -240(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -112(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -236(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -104(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -232(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -96(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -228(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -88(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -224(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -80(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -220(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -72(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -216(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -64(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -212(%ebp), %eax
|
|
Packit |
6c4009 |
vmovsd -56(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
addl $280, %esp
|
|
Packit |
6c4009 |
popq %rbx
|
|
Packit |
6c4009 |
popq %r10
|
|
Packit |
6c4009 |
.cfi_def_cfa 10, 0
|
|
Packit |
6c4009 |
popq %r12
|
|
Packit |
6c4009 |
popq %rbp
|
|
Packit |
6c4009 |
leal -8(%r10), %esp
|
|
Packit |
6c4009 |
.cfi_def_cfa 7, 8
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
.endm
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ENTRY (_ZGVeN8vvv_sincos)
|
|
Packit |
6c4009 |
WRAPPER_IMPL_AVX512_fFF_vvv _ZGVdN4vl8l8_sincos
|
|
Packit |
6c4009 |
END (_ZGVeN8vvv_sincos)
|