1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132
|
/* Double-precision vector (SVE) exp10 function.
Copyright (C) 2023-2025 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library; if not, see
<https://www.gnu.org/licenses/>. */
#include "sv_math.h"
#define SpecialBound 307.0 /* floor (log10 (2^1023)). */
static const struct data
{
double c1, c3, c2, c4, c0;
double shift, log10_2, log2_10_hi, log2_10_lo, scale_thres, special_bound;
} data = {
/* Coefficients generated using Remez algorithm.
rel error: 0x1.9fcb9b3p-60
abs error: 0x1.a20d9598p-60 in [ -log10(2)/128, log10(2)/128 ]
max ulp err 0.52 +0.5. */
.c0 = 0x1.26bb1bbb55516p1,
.c1 = 0x1.53524c73cd32ap1,
.c2 = 0x1.0470591daeafbp1,
.c3 = 0x1.2bd77b1361ef6p0,
.c4 = 0x1.142b5d54e9621p-1,
/* 1.5*2^46+1023. This value is further explained below. */
.shift = 0x1.800000000ffc0p+46,
.log10_2 = 0x1.a934f0979a371p1, /* 1/log2(10). */
.log2_10_hi = 0x1.34413509f79ffp-2, /* log2(10). */
.log2_10_lo = -0x1.9dc1da994fd21p-59,
.scale_thres = 1280.0,
.special_bound = SpecialBound,
};
#define SpecialOffset 0x6000000000000000 /* 0x1p513. */
/* SpecialBias1 + SpecialBias1 = asuint(1.0). */
#define SpecialBias1 0x7000000000000000 /* 0x1p769. */
#define SpecialBias2 0x3010000000000000 /* 0x1p-254. */
/* Update of both special and non-special cases, if any special case is
detected. */
static inline svfloat64_t
special_case (svbool_t pg, svfloat64_t s, svfloat64_t y, svfloat64_t n,
const struct data *d)
{
/* s=2^n may overflow, break it up into s=s1*s2,
such that exp = s + s*y can be computed as s1*(s2+s2*y)
and s1*s1 overflows only if n>0. */
/* If n<=0 then set b to 0x6, 0 otherwise. */
svbool_t p_sign = svcmple (pg, n, 0.0); /* n <= 0. */
svuint64_t b = svdup_u64_z (p_sign, SpecialOffset);
/* Set s1 to generate overflow depending on sign of exponent n. */
svfloat64_t s1 = svreinterpret_f64 (svsubr_x (pg, b, SpecialBias1));
/* Offset s to avoid overflow in final result if n is below threshold. */
svfloat64_t s2 = svreinterpret_f64 (
svadd_x (pg, svsub_x (pg, svreinterpret_u64 (s), SpecialBias2), b));
/* |n| > 1280 => 2^(n) overflows. */
svbool_t p_cmp = svacgt (pg, n, d->scale_thres);
svfloat64_t r1 = svmul_x (svptrue_b64 (), s1, s1);
svfloat64_t r2 = svmla_x (pg, s2, s2, y);
svfloat64_t r0 = svmul_x (svptrue_b64 (), r2, s1);
return svsel (p_cmp, r1, r0);
}
/* Fast vector implementation of exp10 using FEXPA instruction.
Maximum measured error is 1.02 ulp.
SV_NAME_D1 (exp10)(-0x1.2862fec805e58p+2) got 0x1.885a89551d782p-16
want 0x1.885a89551d781p-16. */
svfloat64_t SV_NAME_D1 (exp10) (svfloat64_t x, svbool_t pg)
{
const struct data *d = ptr_barrier (&data);
svbool_t no_big_scale = svacle (pg, x, d->special_bound);
svbool_t special = svnot_z (pg, no_big_scale);
/* n = round(x/(log10(2)/N)). */
svfloat64_t shift = sv_f64 (d->shift);
svfloat64_t z = svmla_x (pg, shift, x, d->log10_2);
svfloat64_t n = svsub_x (pg, z, shift);
/* r = x - n*log10(2)/N. */
svfloat64_t log2_10 = svld1rq (svptrue_b64 (), &d->log2_10_hi);
svfloat64_t r = x;
r = svmls_lane (r, n, log2_10, 0);
r = svmls_lane (r, n, log2_10, 1);
/* scale = 2^(n/N), computed using FEXPA. FEXPA does not propagate NaNs, so
for consistent NaN handling we have to manually propagate them. This
comes at significant performance cost. */
svuint64_t u = svreinterpret_u64 (z);
svfloat64_t scale = svexpa (u);
svfloat64_t c24 = svld1rq (svptrue_b64 (), &d->c2);
/* Approximate exp10(r) using polynomial. */
svfloat64_t r2 = svmul_x (svptrue_b64 (), r, r);
svfloat64_t p12 = svmla_lane (sv_f64 (d->c1), r, c24, 0);
svfloat64_t p34 = svmla_lane (sv_f64 (d->c3), r, c24, 1);
svfloat64_t p14 = svmla_x (pg, p12, p34, r2);
svfloat64_t y = svmla_x (pg, svmul_x (svptrue_b64 (), r, d->c0), r2, p14);
/* Assemble result as exp10(x) = 2^n * exp10(r). If |x| > SpecialBound
multiplication may overflow, so use special case routine. */
if (__glibc_unlikely (svptest_any (pg, special)))
{
/* FEXPA zeroes the sign bit, however the sign is meaningful to the
special case function so needs to be copied.
e = sign bit of u << 46. */
svuint64_t e = svand_x (pg, svlsl_x (pg, u, 46), 0x8000000000000000);
/* Copy sign to scale. */
scale = svreinterpret_f64 (svadd_x (pg, e, svreinterpret_u64 (scale)));
return special_case (pg, scale, y, n, d);
}
/* No special case. */
return svmla_x (pg, scale, scale, y);
}
|