1 | /* Double-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 | #define P(i) sv_f64 (__v_log_data.poly[i]) |
23 | #define N (1 << V_LOG_TABLE_BITS) |
24 | #define Off (0x3fe6900900000000) |
25 | #define MaxTop (0x7ff) |
26 | #define MinTop (0x001) |
27 | #define ThreshTop (0x7fe) /* MaxTop - MinTop. */ |
28 | |
29 | static svfloat64_t NOINLINE |
30 | special_case (svfloat64_t x, svfloat64_t y, svbool_t cmp) |
31 | { |
32 | return sv_call_f64 (f: log, x, y, cmp); |
33 | } |
34 | |
35 | /* SVE port of AdvSIMD log algorithm. |
36 | Maximum measured error is 2.17 ulp: |
37 | SV_NAME_D1 (log)(0x1.a6129884398a3p+0) got 0x1.ffffff1cca043p-2 |
38 | want 0x1.ffffff1cca045p-2. */ |
39 | svfloat64_t SV_NAME_D1 (log) (svfloat64_t x, const svbool_t pg) |
40 | { |
41 | svuint64_t ix = svreinterpret_u64 (x); |
42 | svuint64_t top = svlsr_x (pg, ix, 52); |
43 | svbool_t cmp = svcmpge (pg, svsub_x (pg, top, MinTop), sv_u64 (ThreshTop)); |
44 | |
45 | /* x = 2^k z; where z is in range [Off,2*Off) and exact. |
46 | The range is split into N subintervals. |
47 | The ith subinterval contains z and c is near its center. */ |
48 | svuint64_t tmp = svsub_x (pg, ix, Off); |
49 | /* Calculate table index = (tmp >> (52 - V_LOG_TABLE_BITS)) % N. |
50 | The actual value of i is double this due to table layout. */ |
51 | svuint64_t i |
52 | = svand_x (pg, svlsr_x (pg, tmp, (51 - V_LOG_TABLE_BITS)), (N - 1) << 1); |
53 | svint64_t k |
54 | = svasr_x (pg, svreinterpret_s64 (tmp), 52); /* Arithmetic shift. */ |
55 | svuint64_t iz = svsub_x (pg, ix, svand_x (pg, tmp, 0xfffULL << 52)); |
56 | svfloat64_t z = svreinterpret_f64 (iz); |
57 | /* Lookup in 2 global lists (length N). */ |
58 | svfloat64_t invc = svld1_gather_index (pg, &__v_log_data.table[0].invc, i); |
59 | svfloat64_t logc = svld1_gather_index (pg, &__v_log_data.table[0].logc, i); |
60 | |
61 | /* log(x) = log1p(z/c-1) + log(c) + k*Ln2. */ |
62 | svfloat64_t r = svmad_x (pg, invc, z, -1); |
63 | svfloat64_t kd = svcvt_f64_x (pg, k); |
64 | /* hi = r + log(c) + k*Ln2. */ |
65 | svfloat64_t hi = svmla_x (pg, svadd_x (pg, logc, r), kd, __v_log_data.ln2); |
66 | /* y = r2*(A0 + r*A1 + r2*(A2 + r*A3 + r2*A4)) + hi. */ |
67 | svfloat64_t r2 = svmul_x (pg, r, r); |
68 | svfloat64_t y = svmla_x (pg, P (2), r, P (3)); |
69 | svfloat64_t p = svmla_x (pg, P (0), r, P (1)); |
70 | y = svmla_x (pg, y, r2, P (4)); |
71 | y = svmla_x (pg, p, r2, y); |
72 | |
73 | if (__glibc_unlikely (svptest_any (pg, cmp))) |
74 | return special_case (x, y: svmla_x (svnot_z (pg, cmp), hi, r2, y), cmp); |
75 | return svmla_x (pg, hi, r2, y); |
76 | } |
77 | |