sysdeps/x86_64/fpu/multiarch/svml_d_exp4_core_avx2.S

   1 /* Function exp vectorized with AVX2.
   2    Copyright (C) 2014-2019 Free Software Foundation, Inc.
   3    This file is part of the GNU C Library.
   4
   5    The GNU C Library is free software; you can redistribute it and/or
   6    modify it under the terms of the GNU Lesser General Public
   7    License as published by the Free Software Foundation; either
   8    version 2.1 of the License, or (at your option) any later version.
   9
  10    The GNU C Library is distributed in the hope that it will be useful,
  11    but WITHOUT ANY WARRANTY; without even the implied warranty of
  12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13    Lesser General Public License for more details.
  14
  15    You should have received a copy of the GNU Lesser General Public
  16    License along with the GNU C Library; if not, see
  17    <https://www.gnu.org/licenses/>.  */
  18
  19 #include <sysdep.h>
  20 #include "svml_d_exp_data.h"
  21
  22         .text
  23 ENTRY (_ZGVdN4v_exp_avx2)
  24 /*
  25    ALGORITHM DESCRIPTION:
  26
  27      Argument representation:
  28      N = rint(X*2^k/ln2) = 2^k*M+j
  29      X = N*ln2/2^k + r = M*ln2 + ln2*(j/2^k) + r
  30      then -ln2/2^(k+1) < r < ln2/2^(k+1)
  31      Alternatively:
  32      N = trunc(X*2^k/ln2)
  33      then 0 < r < ln2/2^k
  34
  35      Result calculation:
  36      exp(X) = exp(M*ln2 + ln2*(j/2^k) + r)
  37      = 2^M * 2^(j/2^k) * exp(r)
  38      2^M is calculated by bit manipulation
  39      2^(j/2^k) is stored in table
  40      exp(r) is approximated by polynomial
  41
  42      The table lookup is skipped if k = 0.  */
  43
  44         pushq     %rbp
  45         cfi_adjust_cfa_offset (8)
  46         cfi_rel_offset (%rbp, 0)
  47         movq      %rsp, %rbp
  48         cfi_def_cfa_register (%rbp)
  49         andq      $-64, %rsp
  50         subq      $448, %rsp
  51         movq      __svml_dexp_data@GOTPCREL(%rip), %rax
  52         vmovdqa   %ymm0, %ymm2
  53         vmovupd __dbInvLn2(%rax), %ymm3
  54         vmovupd __dbShifter(%rax), %ymm1
  55         vmovupd __lIndexMask(%rax), %ymm4
  56
  57 /* dM = X*dbInvLn2+dbShifter, dbInvLn2 = 2^k/Ln2 */
  58         vfmadd213pd %ymm1, %ymm2, %ymm3
  59
  60 /* iAbsX = (int)(lX>>32), lX = *(longlong*)&X */
  61         vextracti128 $1, %ymm2, %xmm5
  62         vshufps   $221, %xmm5, %xmm2, %xmm6
  63
  64 /* iAbsX = iAbsX&iAbsMask */
  65         vandps __iAbsMask(%rax), %xmm6, %xmm7
  66
  67 /* dN = dM-dbShifter, dN = rint(X*2^k/Ln2) */
  68         vsubpd    %ymm1, %ymm3, %ymm6
  69
  70 /* iRangeMask = (iAbsX>iDomainRange) */
  71         vpcmpgtd __iDomainRange(%rax), %xmm7, %xmm0
  72         vmovupd __dbLn2hi(%rax), %ymm1
  73         vmovupd __dPC0(%rax), %ymm7
  74
  75 /* Mask = iRangeMask?1:0, set mask for overflow/underflow */
  76         vmovmskps %xmm0, %ecx
  77         vmovupd __dPC2(%rax), %ymm0
  78
  79 /* dR = X - dN*dbLn2hi, dbLn2hi is 52-8-k hi bits of ln2/2^k */
  80         vmovdqa   %ymm2, %ymm5
  81         vfnmadd231pd %ymm6, %ymm1, %ymm5
  82
  83 /* dR = dR - dN*dbLn2lo, dbLn2lo is 40..94 bits of lo part of ln2/2^k */
  84         vfnmadd132pd __dbLn2lo(%rax), %ymm5, %ymm6
  85
  86 /* exp(r) = b0+r*(b0+r*(b1+r*b2)) */
  87         vfmadd213pd __dPC1(%rax), %ymm6, %ymm0
  88         vfmadd213pd %ymm7, %ymm6, %ymm0
  89         vfmadd213pd %ymm7, %ymm6, %ymm0
  90
  91 /* lIndex = (*(longlong*)&dM)&lIndexMask, lIndex is the lower K bits of lM */
  92         vandps    %ymm4, %ymm3, %ymm1
  93
  94 /* table lookup for dT[j] = 2^(j/2^k) */
  95         vxorpd    %ymm6, %ymm6, %ymm6
  96         vpcmpeqd  %ymm5, %ymm5, %ymm5
  97         vgatherqpd %ymm5, (%rax,%ymm1,8), %ymm6
  98
  99 /* lM = (*(longlong*)&dM)&(~lIndexMask) */
 100         vpandn    %ymm3, %ymm4, %ymm3
 101
 102 /* 2^(j/2^k) * exp(r) */
 103         vmulpd    %ymm0, %ymm6, %ymm0
 104
 105 /* lM = lM<<(52-K), 2^M */
 106         vpsllq    $42, %ymm3, %ymm4
 107
 108 /* multiply by 2^M through integer add */
 109         vpaddq    %ymm4, %ymm0, %ymm0
 110         testl     %ecx, %ecx
 111         jne       .LBL_1_3
 112
 113 .LBL_1_2:
 114         cfi_remember_state
 115         movq      %rbp, %rsp
 116         cfi_def_cfa_register (%rsp)
 117         popq      %rbp
 118         cfi_adjust_cfa_offset (-8)
 119         cfi_restore (%rbp)
 120         ret
 121
 122 .LBL_1_3:
 123         cfi_restore_state
 124         vmovupd   %ymm2, 320(%rsp)
 125         vmovupd   %ymm0, 384(%rsp)
 126         je        .LBL_1_2
 127
 128         xorb      %dl, %dl
 129         xorl      %eax, %eax
 130         vmovups   %ymm8, 224(%rsp)
 131         vmovups   %ymm9, 192(%rsp)
 132         vmovups   %ymm10, 160(%rsp)
 133         vmovups   %ymm11, 128(%rsp)
 134         vmovups   %ymm12, 96(%rsp)
 135         vmovups   %ymm13, 64(%rsp)
 136         vmovups   %ymm14, 32(%rsp)
 137         vmovups   %ymm15, (%rsp)
 138         movq      %rsi, 264(%rsp)
 139         movq      %rdi, 256(%rsp)
 140         movq      %r12, 296(%rsp)
 141         cfi_offset_rel_rsp (12, 296)
 142         movb      %dl, %r12b
 143         movq      %r13, 288(%rsp)
 144         cfi_offset_rel_rsp (13, 288)
 145         movl      %ecx, %r13d
 146         movq      %r14, 280(%rsp)
 147         cfi_offset_rel_rsp (14, 280)
 148         movl      %eax, %r14d
 149         movq      %r15, 272(%rsp)
 150         cfi_offset_rel_rsp (15, 272)
 151         cfi_remember_state
 152
 153 .LBL_1_6:
 154         btl       %r14d, %r13d
 155         jc        .LBL_1_12
 156
 157 .LBL_1_7:
 158         lea       1(%r14), %esi
 159         btl       %esi, %r13d
 160         jc        .LBL_1_10
 161
 162 .LBL_1_8:
 163         incb      %r12b
 164         addl      $2, %r14d
 165         cmpb      $16, %r12b
 166         jb        .LBL_1_6
 167
 168         vmovups   224(%rsp), %ymm8
 169         vmovups   192(%rsp), %ymm9
 170         vmovups   160(%rsp), %ymm10
 171         vmovups   128(%rsp), %ymm11
 172         vmovups   96(%rsp), %ymm12
 173         vmovups   64(%rsp), %ymm13
 174         vmovups   32(%rsp), %ymm14
 175         vmovups   (%rsp), %ymm15
 176         vmovupd   384(%rsp), %ymm0
 177         movq      264(%rsp), %rsi
 178         movq      256(%rsp), %rdi
 179         movq      296(%rsp), %r12
 180         cfi_restore (%r12)
 181         movq      288(%rsp), %r13
 182         cfi_restore (%r13)
 183         movq      280(%rsp), %r14
 184         cfi_restore (%r14)
 185         movq      272(%rsp), %r15
 186         cfi_restore (%r15)
 187         jmp       .LBL_1_2
 188
 189 .LBL_1_10:
 190         cfi_restore_state
 191         movzbl    %r12b, %r15d
 192         shlq      $4, %r15
 193         vmovsd    328(%rsp,%r15), %xmm0
 194         vzeroupper
 195
 196         call      JUMPTARGET(__exp_finite)
 197
 198         vmovsd    %xmm0, 392(%rsp,%r15)
 199         jmp       .LBL_1_8
 200
 201 .LBL_1_12:
 202         movzbl    %r12b, %r15d
 203         shlq      $4, %r15
 204         vmovsd    320(%rsp,%r15), %xmm0
 205         vzeroupper
 206
 207         call      JUMPTARGET(__exp_finite)
 208
 209         vmovsd    %xmm0, 384(%rsp,%r15)
 210         jmp       .LBL_1_7
 211
 212 END (_ZGVdN4v_exp_avx2)