1 | /* Double-precision vector (SVE) exp2 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_EXP_TABLE_BITS) |
24 | |
25 | #define BigBound 1022 |
26 | #define UOFlowBound 1280 |
27 | |
28 | static const struct data |
29 | { |
30 | double poly[4]; |
31 | double shift, big_bound, uoflow_bound; |
32 | } data = { |
33 | /* Coefficients are computed using Remez algorithm with |
34 | minimisation of the absolute error. */ |
35 | .poly = { 0x1.62e42fefa3686p-1, 0x1.ebfbdff82c241p-3, 0x1.c6b09b16de99ap-5, |
36 | 0x1.3b2abf5571ad8p-7 }, |
37 | .shift = 0x1.8p52 / N, |
38 | .uoflow_bound = UOFlowBound, |
39 | .big_bound = BigBound, |
40 | }; |
41 | |
42 | #define SpecialOffset 0x6000000000000000 /* 0x1p513. */ |
43 | /* SpecialBias1 + SpecialBias1 = asuint(1.0). */ |
44 | #define SpecialBias1 0x7000000000000000 /* 0x1p769. */ |
45 | #define SpecialBias2 0x3010000000000000 /* 0x1p-254. */ |
46 | |
47 | /* Update of both special and non-special cases, if any special case is |
48 | detected. */ |
49 | static inline svfloat64_t |
50 | special_case (svbool_t pg, svfloat64_t s, svfloat64_t y, svfloat64_t n, |
51 | const struct data *d) |
52 | { |
53 | /* s=2^n may overflow, break it up into s=s1*s2, |
54 | such that exp = s + s*y can be computed as s1*(s2+s2*y) |
55 | and s1*s1 overflows only if n>0. */ |
56 | |
57 | /* If n<=0 then set b to 0x6, 0 otherwise. */ |
58 | svbool_t p_sign = svcmple (pg, n, 0.0); /* n <= 0. */ |
59 | svuint64_t b = svdup_u64_z (p_sign, SpecialOffset); |
60 | |
61 | /* Set s1 to generate overflow depending on sign of exponent n. */ |
62 | svfloat64_t s1 = svreinterpret_f64 (svsubr_x (pg, b, SpecialBias1)); |
63 | /* Offset s to avoid overflow in final result if n is below threshold. */ |
64 | svfloat64_t s2 = svreinterpret_f64 ( |
65 | svadd_x (pg, svsub_x (pg, svreinterpret_u64 (s), SpecialBias2), b)); |
66 | |
67 | /* |n| > 1280 => 2^(n) overflows. */ |
68 | svbool_t p_cmp = svacgt (pg, n, d->uoflow_bound); |
69 | |
70 | svfloat64_t r1 = svmul_x (pg, s1, s1); |
71 | svfloat64_t r2 = svmla_x (pg, s2, s2, y); |
72 | svfloat64_t r0 = svmul_x (pg, r2, s1); |
73 | |
74 | return svsel (p_cmp, r1, r0); |
75 | } |
76 | |
77 | /* Fast vector implementation of exp2. |
78 | Maximum measured error is 1.65 ulp. |
79 | _ZGVsMxv_exp2(-0x1.4c264ab5b559bp-6) got 0x1.f8db0d4df721fp-1 |
80 | want 0x1.f8db0d4df721dp-1. */ |
81 | svfloat64_t SV_NAME_D1 (exp2) (svfloat64_t x, svbool_t pg) |
82 | { |
83 | const struct data *d = ptr_barrier (&data); |
84 | svbool_t no_big_scale = svacle (pg, x, d->big_bound); |
85 | svbool_t special = svnot_z (pg, no_big_scale); |
86 | |
87 | /* Reduce x to k/N + r, where k is integer and r in [-1/2N, 1/2N]. */ |
88 | svfloat64_t shift = sv_f64 (x: d->shift); |
89 | svfloat64_t kd = svadd_x (pg, x, shift); |
90 | svuint64_t ki = svreinterpret_u64 (kd); |
91 | /* kd = k/N. */ |
92 | kd = svsub_x (pg, kd, shift); |
93 | svfloat64_t r = svsub_x (pg, x, kd); |
94 | |
95 | /* scale ~= 2^(k/N). */ |
96 | svuint64_t idx = svand_x (pg, ki, N - 1); |
97 | svuint64_t sbits = svld1_gather_index (pg, __v_exp_data, idx); |
98 | /* This is only a valid scale when -1023*N < k < 1024*N. */ |
99 | svuint64_t top = svlsl_x (pg, ki, 52 - V_EXP_TABLE_BITS); |
100 | svfloat64_t scale = svreinterpret_f64 (svadd_x (pg, sbits, top)); |
101 | |
102 | /* Approximate exp2(r) using polynomial. */ |
103 | svfloat64_t r2 = svmul_x (pg, r, r); |
104 | svfloat64_t p = sv_pairwise_poly_3_f64_x (pg, x: r, x2: r2, poly: d->poly); |
105 | svfloat64_t y = svmul_x (pg, r, p); |
106 | |
107 | /* Assemble exp2(x) = exp2(r) * scale. */ |
108 | if (__glibc_unlikely (svptest_any (pg, special))) |
109 | return special_case (pg, s: scale, y, n: kd, d); |
110 | return svmla_x (pg, scale, scale, y); |
111 | } |
112 | |