1 | /* Single-precision vector (SVE) log2 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 | |
22 | static const struct data |
23 | { |
24 | float poly_02468[5]; |
25 | float poly_1357[4]; |
26 | } data = { |
27 | .poly_1357 = { |
28 | /* Coefficients copied from the AdvSIMD routine, then rearranged so that coeffs |
29 | 1, 3, 5 and 7 can be loaded as a single quad-word, hence used with _lane |
30 | variant of MLA intrinsic. */ |
31 | -0x1.715458p-1f, -0x1.7171a4p-2f, -0x1.e5143ep-3f, -0x1.c675bp-3f |
32 | }, |
33 | .poly_02468 = { 0x1.715476p0f, 0x1.ec701cp-2f, 0x1.27a0b8p-2f, |
34 | 0x1.9d8ecap-3f, 0x1.9e495p-3f }, |
35 | }; |
36 | |
37 | #define Min (0x00800000) |
38 | #define Max (0x7f800000) |
39 | #define Thres (0x7f000000) /* Max - Min. */ |
40 | #define MantissaMask (0x007fffff) |
41 | #define Off (0x3f2aaaab) /* 0.666667. */ |
42 | |
43 | static svfloat32_t NOINLINE |
44 | special_case (svfloat32_t x, svfloat32_t y, svbool_t cmp) |
45 | { |
46 | return sv_call_f32 (f: log2f, x, y, cmp); |
47 | } |
48 | |
49 | /* Optimised implementation of SVE log2f, using the same algorithm |
50 | and polynomial as AdvSIMD log2f. |
51 | Maximum error is 2.48 ULPs: |
52 | SV_NAME_F1 (log2)(0x1.558174p+0) got 0x1.a9be84p-2 |
53 | want 0x1.a9be8p-2. */ |
54 | svfloat32_t SV_NAME_F1 (log2) (svfloat32_t x, const svbool_t pg) |
55 | { |
56 | const struct data *d = ptr_barrier (&data); |
57 | |
58 | svuint32_t u = svreinterpret_u32 (x); |
59 | svbool_t special = svcmpge (pg, svsub_x (pg, u, Min), Thres); |
60 | |
61 | /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3. */ |
62 | u = svsub_x (pg, u, Off); |
63 | svfloat32_t n = svcvt_f32_x ( |
64 | pg, svasr_x (pg, svreinterpret_s32 (u), 23)); /* Sign-extend. */ |
65 | u = svand_x (pg, u, MantissaMask); |
66 | u = svadd_x (pg, u, Off); |
67 | svfloat32_t r = svsub_x (pg, svreinterpret_f32 (u), 1.0f); |
68 | |
69 | /* y = log2(1+r) + n. */ |
70 | svfloat32_t r2 = svmul_x (pg, r, r); |
71 | |
72 | /* Evaluate polynomial using pairwise Horner scheme. */ |
73 | svfloat32_t p_1357 = svld1rq (svptrue_b32 (), &d->poly_1357[0]); |
74 | svfloat32_t q_01 = svmla_lane (sv_f32 (x: d->poly_02468[0]), r, p_1357, 0); |
75 | svfloat32_t q_23 = svmla_lane (sv_f32 (x: d->poly_02468[1]), r, p_1357, 1); |
76 | svfloat32_t q_45 = svmla_lane (sv_f32 (x: d->poly_02468[2]), r, p_1357, 2); |
77 | svfloat32_t q_67 = svmla_lane (sv_f32 (x: d->poly_02468[3]), r, p_1357, 3); |
78 | svfloat32_t y = svmla_x (pg, q_67, r2, sv_f32 (x: d->poly_02468[4])); |
79 | y = svmla_x (pg, q_45, r2, y); |
80 | y = svmla_x (pg, q_23, r2, y); |
81 | y = svmla_x (pg, q_01, r2, y); |
82 | |
83 | if (__glibc_unlikely (svptest_any (pg, special))) |
84 | return special_case (x, y: svmla_x (svnot_z (pg, special), n, r, y), cmp: special); |
85 | return svmla_x (pg, n, r, y); |
86 | } |
87 | |