1 | /* Single-precision vector (SVE) log 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_0135[4]; |
25 | float poly_246[3]; |
26 | float ln2; |
27 | } data = { |
28 | .poly_0135 = { |
29 | /* Coefficients copied from the AdvSIMD routine in math/, then rearranged so |
30 | that coeffs 0, 1, 3 and 5 can be loaded as a single quad-word, hence used |
31 | with _lane variant of MLA intrinsic. */ |
32 | -0x1.3e737cp-3f, 0x1.5a9aa2p-3f, 0x1.961348p-3f, 0x1.555d7cp-2f |
33 | }, |
34 | .poly_246 = { -0x1.4f9934p-3f, -0x1.00187cp-2f, -0x1.ffffc8p-2f }, |
35 | .ln2 = 0x1.62e43p-1f |
36 | }; |
37 | |
38 | #define Min (0x00800000) |
39 | #define Max (0x7f800000) |
40 | #define Thresh (0x7f000000) /* Max - Min. */ |
41 | #define Mask (0x007fffff) |
42 | #define Off (0x3f2aaaab) /* 0.666667. */ |
43 | |
44 | static svfloat32_t NOINLINE |
45 | special_case (svfloat32_t x, svfloat32_t y, svbool_t cmp) |
46 | { |
47 | return sv_call_f32 (f: logf, x, y, cmp); |
48 | } |
49 | |
50 | /* Optimised implementation of SVE logf, using the same algorithm and |
51 | polynomial as the AdvSIMD routine. Maximum error is 3.34 ULPs: |
52 | SV_NAME_F1 (log)(0x1.557298p+0) got 0x1.26edecp-2 |
53 | want 0x1.26ede6p-2. */ |
54 | svfloat32_t SV_NAME_F1 (log) (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 cmp = svcmpge (pg, svsub_x (pg, u, Min), Thresh); |
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, Mask); |
66 | u = svadd_x (pg, u, Off); |
67 | svfloat32_t r = svsub_x (pg, svreinterpret_f32 (u), 1.0f); |
68 | |
69 | /* y = log(1+r) + n*ln2. */ |
70 | svfloat32_t r2 = svmul_x (pg, r, r); |
71 | /* n*ln2 + r + r2*(P6 + r*P5 + r2*(P4 + r*P3 + r2*(P2 + r*P1 + r2*P0))). */ |
72 | svfloat32_t p_0135 = svld1rq (svptrue_b32 (), &d->poly_0135[0]); |
73 | svfloat32_t p = svmla_lane (sv_f32 (x: d->poly_246[0]), r, p_0135, 1); |
74 | svfloat32_t q = svmla_lane (sv_f32 (x: d->poly_246[1]), r, p_0135, 2); |
75 | svfloat32_t y = svmla_lane (sv_f32 (x: d->poly_246[2]), r, p_0135, 3); |
76 | p = svmla_lane (p, r2, p_0135, 0); |
77 | |
78 | q = svmla_x (pg, q, r2, p); |
79 | y = svmla_x (pg, y, r2, q); |
80 | p = svmla_x (pg, r, n, d->ln2); |
81 | |
82 | if (__glibc_unlikely (svptest_any (pg, cmp))) |
83 | return special_case (x, y: svmla_x (svnot_z (pg, cmp), p, r2, y), cmp); |
84 | return svmla_x (pg, p, r2, y); |
85 | } |
86 | |