sysdeps/aarch64/fpu/log2_sve.c

   1 /* Double-precision vector (SVE) log2 function
   2
   3    Copyright (C) 2023-2024 Free Software Foundation, Inc.
   4    This file is part of the GNU C Library.
   5
   6    The GNU C Library is free software; you can redistribute it and/or
   7    modify it under the terms of the GNU Lesser General Public
   8    License as published by the Free Software Foundation; either
   9    version 2.1 of the License, or (at your option) any later version.
  10
  11    The GNU C Library is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14    Lesser General Public License for more details.
  15
  16    You should have received a copy of the GNU Lesser General Public
  17    License along with the GNU C Library; if not, see
  18    <https://www.gnu.org/licenses/>.  */
  19
  20 #include "sv_math.h"
  21 #include "poly_sve_f64.h"
  22
  23 #define N (1 << V_LOG2_TABLE_BITS)
  24 #define Off 0x3fe6900900000000
  25 #define Max (0x7ff0000000000000)
  26 #define Min (0x0010000000000000)
  27 #define Thresh (0x7fe0000000000000) /* Max - Min.  */
  28
  29 static svfloat64_t NOINLINE
  30 special_case (svfloat64_t x, svfloat64_t y, svbool_t cmp)
  31 {
  32   return sv_call_f64 (log2, x, y, cmp);
  33 }
  34
  35 /* Double-precision SVE log2 routine.
  36    Implements the same algorithm as AdvSIMD log10, with coefficients and table
  37    entries scaled in extended precision.
  38    The maximum observed error is 2.58 ULP:
  39    SV_NAME_D1 (log2)(0x1.0b556b093869bp+0) got 0x1.fffb34198d9dap-5
  40                                           want 0x1.fffb34198d9ddp-5.  */
  41 svfloat64_t SV_NAME_D1 (log2) (svfloat64_t x, const svbool_t pg)
  42 {
  43   svuint64_t ix = svreinterpret_u64 (x);
  44   svbool_t special = svcmpge (pg, svsub_x (pg, ix, Min), Thresh);
  45
  46   /* x = 2^k z; where z is in range [Off,2*Off) and exact.
  47      The range is split into N subintervals.
  48      The ith subinterval contains z and c is near its center.  */
  49   svuint64_t tmp = svsub_x (pg, ix, Off);
  50   svuint64_t i = svlsr_x (pg, tmp, 51 - V_LOG2_TABLE_BITS);
  51   i = svand_x (pg, i, (N - 1) << 1);
  52   svfloat64_t k = svcvt_f64_x (pg, svasr_x (pg, svreinterpret_s64 (tmp), 52));
  53   svfloat64_t z = svreinterpret_f64 (
  54       svsub_x (pg, ix, svand_x (pg, tmp, 0xfffULL << 52)));
  55
  56   svfloat64_t invc = svld1_gather_index (pg, &__v_log2_data.table[0].invc, i);
  57   svfloat64_t log2c
  58       = svld1_gather_index (pg, &__v_log2_data.table[0].log2c, i);
  59
  60   /* log2(x) = log1p(z/c-1)/log(2) + log2(c) + k.  */
  61
  62   svfloat64_t r = svmad_x (pg, invc, z, -1.0);
  63   svfloat64_t w = svmla_x (pg, log2c, r, __v_log2_data.invln2);
  64
  65   svfloat64_t r2 = svmul_x (pg, r, r);
  66   svfloat64_t y = sv_pw_horner_4_f64_x (pg, r, r2, __v_log2_data.poly);
  67   w = svadd_x (pg, k, w);
  68
  69   if (__glibc_unlikely (svptest_any (pg, special)))
  70     return special_case (x, svmla_x (svnot_z (pg, special), w, r2, y),
  71                          special);
  72   return svmla_x (pg, w, r2, y);
  73 }