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
|
/* Helper for single-precision AdvSIMD routines which depend on exp
Copyright (C) 2024-2025 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/>. */
#ifndef AARCH64_FPU_V_EXPF_INLINE_H
#define AARCH64_FPU_V_EXPF_INLINE_H
#include "v_math.h"
struct v_expf_data
{
float ln2_hi, ln2_lo, c0, c2;
float32x4_t inv_ln2, c1, c3, c4;
/* asuint(1.0f). */
uint32x4_t exponent_bias;
};
/* maxerr: 1.45358 +0.5 ulp. */
#define V_EXPF_DATA \
{ \
.c0 = 0x1.0e4020p-7f, .c1 = V4 (0x1.573e2ep-5f), .c2 = 0x1.555e66p-3f, \
.c3 = V4 (0x1.fffdb6p-2f), .c4 = V4 (0x1.ffffecp-1f), \
.ln2_hi = 0x1.62e4p-1f, .ln2_lo = 0x1.7f7d1cp-20f, \
.inv_ln2 = V4 (0x1.715476p+0f), .exponent_bias = V4 (0x3f800000), \
}
static inline float32x4_t
v_expf_inline (float32x4_t x, const struct v_expf_data *d)
{
/* Helper routine for calculating exp(ax).
Copied from v_expf.c, with all special-case handling removed - the
calling routine should handle special values if required. */
/* exp(ax) = 2^n (1 + poly(r)), with 1 + poly(r) in [1/sqrt(2),sqrt(2)]
ax = ln2*n + r, with r in [-ln2/2, ln2/2]. */
float32x4_t ax = vabsq_f32 (x);
float32x4_t ln2_c02 = vld1q_f32 (&d->ln2_hi);
float32x4_t n = vrndaq_f32 (vmulq_f32 (ax, d->inv_ln2));
float32x4_t r = vfmsq_laneq_f32 (ax, n, ln2_c02, 0);
r = vfmsq_laneq_f32 (r, n, ln2_c02, 1);
uint32x4_t e = vshlq_n_u32 (vreinterpretq_u32_s32 (vcvtq_s32_f32 (n)), 23);
float32x4_t scale = vreinterpretq_f32_u32 (vaddq_u32 (e, d->exponent_bias));
/* Custom order-4 Estrin avoids building high order monomial. */
float32x4_t r2 = vmulq_f32 (r, r);
float32x4_t p = vfmaq_laneq_f32 (d->c1, r, ln2_c02, 2);
float32x4_t q = vfmaq_laneq_f32 (d->c3, r, ln2_c02, 3);
q = vfmaq_f32 (q, p, r2);
p = vmulq_f32 (d->c4, r);
float32x4_t poly = vfmaq_f32 (p, q, r2);
return vfmaq_f32 (scale, poly, scale);
}
#endif
|
