1/* Double-precision AdvSIMD inverse tan
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 "v_math.h"
21#include "poly_advsimd_f64.h"
22
23static const struct data
24{
25 float64x2_t c0, c2, c4, c6, c8, c10, c12, c14, c16, c18;
26 float64x2_t pi_over_2;
27 double c1, c3, c5, c7, c9, c11, c13, c15, c17, c19;
28} data = {
29 /* Coefficients of polynomial P such that atan(x)~x+x*P(x^2) on
30 [2**-1022, 1.0]. */
31 .c0 = V2 (-0x1.5555555555555p-2), .c1 = 0x1.99999999996c1p-3,
32 .c2 = V2 (-0x1.2492492478f88p-3), .c3 = 0x1.c71c71bc3951cp-4,
33 .c4 = V2 (-0x1.745d160a7e368p-4), .c5 = 0x1.3b139b6a88ba1p-4,
34 .c6 = V2 (-0x1.11100ee084227p-4), .c7 = 0x1.e1d0f9696f63bp-5,
35 .c8 = V2 (-0x1.aebfe7b418581p-5), .c9 = 0x1.842dbe9b0d916p-5,
36 .c10 = V2 (-0x1.5d30140ae5e99p-5), .c11 = 0x1.338e31eb2fbbcp-5,
37 .c12 = V2 (-0x1.00e6eece7de8p-5), .c13 = 0x1.860897b29e5efp-6,
38 .c14 = V2 (-0x1.0051381722a59p-6), .c15 = 0x1.14e9dc19a4a4ep-7,
39 .c16 = V2 (-0x1.d0062b42fe3bfp-9), .c17 = 0x1.17739e210171ap-10,
40 .c18 = V2 (-0x1.ab24da7be7402p-13), .c19 = 0x1.358851160a528p-16,
41 .pi_over_2 = V2 (0x1.921fb54442d18p+0),
42};
43
44#define SignMask v_u64 (0x8000000000000000)
45#define TinyBound 0x3e10000000000000 /* asuint64(0x1p-30). */
46#define BigBound 0x4340000000000000 /* asuint64(0x1p53). */
47
48/* Fast implementation of vector atan.
49 Based on atan(x) ~ shift + z + z^3 * P(z^2) with reduction to [0,1] using
50 z=1/x and shift = pi/2. Maximum observed error is 2.27 ulps:
51 _ZGVnN2v_atan (0x1.0005af27c23e9p+0) got 0x1.9225645bdd7c1p-1
52 want 0x1.9225645bdd7c3p-1. */
53float64x2_t VPCS_ATTR V_NAME_D1 (atan) (float64x2_t x)
54{
55 const struct data *d = ptr_barrier (&data);
56 float64x2_t c13 = vld1q_f64 (&d->c1);
57 float64x2_t c57 = vld1q_f64 (&d->c5);
58 float64x2_t c911 = vld1q_f64 (&d->c9);
59 float64x2_t c1315 = vld1q_f64 (&d->c13);
60 float64x2_t c1719 = vld1q_f64 (&d->c17);
61
62 /* Small cases, infs and nans are supported by our approximation technique,
63 but do not set fenv flags correctly. Only trigger special case if we need
64 fenv. */
65 uint64x2_t ix = vreinterpretq_u64_f64 (x);
66 uint64x2_t sign = vandq_u64 (ix, SignMask);
67
68#if WANT_SIMD_EXCEPT
69 uint64x2_t ia12 = vandq_u64 (ix, v_u64 (0x7ff0000000000000));
70 uint64x2_t special = vcgtq_u64 (vsubq_u64 (ia12, v_u64 (TinyBound)),
71 v_u64 (BigBound - TinyBound));
72 /* If any lane is special, fall back to the scalar routine for all lanes. */
73 if (__glibc_unlikely (v_any_u64 (special)))
74 return v_call_f64 (atan, x, v_f64 (0), v_u64 (-1));
75#endif
76
77 /* Argument reduction:
78 y := arctan(x) for x < 1
79 y := pi/2 + arctan(-1/x) for x > 1
80 Hence, use z=-1/a if x>=1, otherwise z=a. */
81 uint64x2_t red = vcagtq_f64 (x, v_f64 (1.0));
82 /* Avoid dependency in abs(x) in division (and comparison). */
83 float64x2_t z = vbslq_f64 (red, vdivq_f64 (v_f64 (1.0), x), x);
84 float64x2_t shift = vreinterpretq_f64_u64 (
85 vandq_u64 (red, vreinterpretq_u64_f64 (d->pi_over_2)));
86 /* Use absolute value only when needed (odd powers of z). */
87 float64x2_t az = vbslq_f64 (
88 SignMask, vreinterpretq_f64_u64 (vandq_u64 (SignMask, red)), z);
89
90 /* Calculate the polynomial approximation.
91 Use split Estrin scheme for P(z^2) with deg(P)=19. Use split instead of
92 full scheme to avoid underflow in x^16.
93 The order 19 polynomial P approximates
94 (atan(sqrt(x))-sqrt(x))/x^(3/2). */
95 float64x2_t z2 = vmulq_f64 (z, z);
96 float64x2_t x2 = vmulq_f64 (z2, z2);
97 float64x2_t x4 = vmulq_f64 (x2, x2);
98 float64x2_t x8 = vmulq_f64 (x4, x4);
99
100 /* estrin_7. */
101 float64x2_t p01 = vfmaq_laneq_f64 (d->c0, z2, c13, 0);
102 float64x2_t p23 = vfmaq_laneq_f64 (d->c2, z2, c13, 1);
103 float64x2_t p03 = vfmaq_f64 (p01, x2, p23);
104
105 float64x2_t p45 = vfmaq_laneq_f64 (d->c4, z2, c57, 0);
106 float64x2_t p67 = vfmaq_laneq_f64 (d->c6, z2, c57, 1);
107 float64x2_t p47 = vfmaq_f64 (p45, x2, p67);
108
109 float64x2_t p07 = vfmaq_f64 (p03, x4, p47);
110
111 /* estrin_11. */
112 float64x2_t p89 = vfmaq_laneq_f64 (d->c8, z2, c911, 0);
113 float64x2_t p1011 = vfmaq_laneq_f64 (d->c10, z2, c911, 1);
114 float64x2_t p811 = vfmaq_f64 (p89, x2, p1011);
115
116 float64x2_t p1213 = vfmaq_laneq_f64 (d->c12, z2, c1315, 0);
117 float64x2_t p1415 = vfmaq_laneq_f64 (d->c14, z2, c1315, 1);
118 float64x2_t p1215 = vfmaq_f64 (p1213, x2, p1415);
119
120 float64x2_t p1617 = vfmaq_laneq_f64 (d->c16, z2, c1719, 0);
121 float64x2_t p1819 = vfmaq_laneq_f64 (d->c18, z2, c1719, 1);
122 float64x2_t p1619 = vfmaq_f64 (p1617, x2, p1819);
123
124 float64x2_t p815 = vfmaq_f64 (p811, x4, p1215);
125 float64x2_t p819 = vfmaq_f64 (p815, x8, p1619);
126
127 float64x2_t y = vfmaq_f64 (p07, p819, x8);
128
129 /* Finalize. y = shift + z + z^3 * P(z^2). */
130 y = vfmaq_f64 (az, y, vmulq_f64 (z2, az));
131 y = vaddq_f64 (y, shift);
132
133 /* y = atan(x) if x>0, -atan(-x) otherwise. */
134 y = vreinterpretq_f64_u64 (veorq_u64 (vreinterpretq_u64_f64 (y), sign));
135 return y;
136}
137

source code of glibc/sysdeps/aarch64/fpu/atan_advsimd.c