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
|
/* Single-precision vector (SVE) exp10 function.
Copyright (C) 2023-2024 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"
/* For x < -Thres, the result is subnormal and not handled correctly by
FEXPA. */
#define Thres 37.9
static const struct data
{
float log2_10_lo, c0, c2, c4;
float c1, c3, log10_2;
float shift, log2_10_hi, thres;
} data = {
/* Coefficients generated using Remez algorithm with minimisation of relative
error.
rel error: 0x1.89dafa3p-24
abs error: 0x1.167d55p-23 in [-log10(2)/2, log10(2)/2]
maxerr: 0.52 +0.5 ulp. */
.c0 = 0x1.26bb16p+1f,
.c1 = 0x1.5350d2p+1f,
.c2 = 0x1.04744ap+1f,
.c3 = 0x1.2d8176p+0f,
.c4 = 0x1.12b41ap-1f,
/* 1.5*2^17 + 127, a shift value suitable for FEXPA. */
.shift = 0x1.803f8p17f,
.log10_2 = 0x1.a934fp+1,
.log2_10_hi = 0x1.344136p-2,
.log2_10_lo = -0x1.ec10cp-27,
.thres = Thres,
};
static inline svfloat32_t
sv_exp10f_inline (svfloat32_t x, const svbool_t pg, const struct data *d)
{
/* exp10(x) = 2^(n/N) * 10^r = 2^n * (1 + poly (r)),
with poly(r) in [1/sqrt(2), sqrt(2)] and
x = r + n * log10(2) / N, with r in [-log10(2)/2N, log10(2)/2N]. */
svfloat32_t lane_consts = svld1rq (svptrue_b32 (), &d->log2_10_lo);
/* n = round(x/(log10(2)/N)). */
svfloat32_t shift = sv_f32 (d->shift);
svfloat32_t z = svmad_x (pg, sv_f32 (d->log10_2), x, shift);
svfloat32_t n = svsub_x (svptrue_b32 (), z, shift);
/* r = x - n*log10(2)/N. */
svfloat32_t r = svmsb_x (pg, sv_f32 (d->log2_10_hi), n, x);
r = svmls_lane (r, n, lane_consts, 0);
svfloat32_t scale = svexpa (svreinterpret_u32 (z));
/* Polynomial evaluation: poly(r) ~ exp10(r)-1. */
svfloat32_t p12 = svmla_lane (sv_f32 (d->c1), r, lane_consts, 2);
svfloat32_t p34 = svmla_lane (sv_f32 (d->c3), r, lane_consts, 3);
svfloat32_t r2 = svmul_x (svptrue_b32 (), r, r);
svfloat32_t p14 = svmla_x (pg, p12, p34, r2);
svfloat32_t p0 = svmul_lane (r, lane_consts, 1);
svfloat32_t poly = svmla_x (pg, p0, r2, p14);
return svmla_x (pg, scale, scale, poly);
}
static svfloat32_t NOINLINE
special_case (svfloat32_t x, svbool_t special, const struct data *d)
{
return sv_call_f32 (exp10f, x, sv_exp10f_inline (x, svptrue_b32 (), d),
special);
}
/* Single-precision SVE exp10f routine. Implements the same algorithm
as AdvSIMD exp10f.
Worst case error is 1.02 ULPs.
_ZGVsMxv_exp10f(-0x1.040488p-4) got 0x1.ba5f9ep-1
want 0x1.ba5f9cp-1. */
svfloat32_t SV_NAME_F1 (exp10) (svfloat32_t x, const svbool_t pg)
{
const struct data *d = ptr_barrier (&data);
svbool_t special = svacgt (pg, x, d->thres);
if (__glibc_unlikely (svptest_any (special, special)))
return special_case (x, special, d);
return sv_exp10f_inline (x, pg, d);
}
|