1 | /* |
2 | * Double-precision vector e^x function. |
3 | * |
4 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
5 | * See https://llvm.org/LICENSE.txt for license information. |
6 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
7 | */ |
8 | |
9 | #include "mathlib.h" |
10 | #include "v_math.h" |
11 | #if V_SUPPORTED |
12 | #include "v_exp.h" |
13 | |
14 | #if V_EXP_TABLE_BITS == 7 |
15 | /* maxerr: 1.88 +0.5 ulp |
16 | rel error: 1.4337*2^-53 |
17 | abs error: 1.4299*2^-53 in [ -ln2/256, ln2/256 ]. */ |
18 | #define C1 v_f64 (0x1.ffffffffffd43p-2) |
19 | #define C2 v_f64 (0x1.55555c75adbb2p-3) |
20 | #define C3 v_f64 (0x1.55555da646206p-5) |
21 | #define InvLn2 v_f64 (0x1.71547652b82fep7) /* N/ln2. */ |
22 | #define Ln2hi v_f64 (0x1.62e42fefa39efp-8) /* ln2/N. */ |
23 | #define Ln2lo v_f64 (0x1.abc9e3b39803f3p-63) |
24 | #elif V_EXP_TABLE_BITS == 8 |
25 | /* maxerr: 0.54 +0.5 ulp |
26 | rel error: 1.4318*2^-58 |
27 | abs error: 1.4299*2^-58 in [ -ln2/512, ln2/512 ]. */ |
28 | #define C1 v_f64 (0x1.fffffffffffd4p-2) |
29 | #define C2 v_f64 (0x1.5555571d6b68cp-3) |
30 | #define C3 v_f64 (0x1.5555576a59599p-5) |
31 | #define InvLn2 v_f64 (0x1.71547652b82fep8) |
32 | #define Ln2hi v_f64 (0x1.62e42fefa39efp-9) |
33 | #define Ln2lo v_f64 (0x1.abc9e3b39803f3p-64) |
34 | #endif |
35 | |
36 | #define N (1 << V_EXP_TABLE_BITS) |
37 | #define Tab __v_exp_data |
38 | #define IndexMask v_u64 (N - 1) |
39 | #define Shift v_f64 (0x1.8p+52) |
40 | #define Thres v_f64 (704.0) |
41 | |
42 | VPCS_ATTR |
43 | static v_f64_t |
44 | specialcase (v_f64_t s, v_f64_t y, v_f64_t n) |
45 | { |
46 | v_f64_t absn = v_abs_f64 (n); |
47 | |
48 | /* 2^(n/N) may overflow, break it up into s1*s2. */ |
49 | v_u64_t b = v_cond_u64 (n <= v_f64 (0.0)) & v_u64 (0x6000000000000000); |
50 | v_f64_t s1 = v_as_f64_u64 (v_u64 (0x7000000000000000) - b); |
51 | v_f64_t s2 = v_as_f64_u64 (v_as_u64_f64 (s) - v_u64 (0x3010000000000000) + b); |
52 | v_u64_t cmp = v_cond_u64 (absn > v_f64 (1280.0 * N)); |
53 | v_f64_t r1 = s1 * s1; |
54 | v_f64_t r0 = v_fma_f64 (y, s2, s2) * s1; |
55 | return v_as_f64_u64 ((cmp & v_as_u64_f64 (r1)) | (~cmp & v_as_u64_f64 (r0))); |
56 | } |
57 | |
58 | VPCS_ATTR |
59 | v_f64_t |
60 | V_NAME(exp) (v_f64_t x) |
61 | { |
62 | v_f64_t n, r, r2, s, y, z; |
63 | v_u64_t cmp, u, e, i; |
64 | |
65 | cmp = v_cond_u64 (v_abs_f64 (x) > Thres); |
66 | |
67 | /* n = round(x/(ln2/N)). */ |
68 | z = v_fma_f64 (x, InvLn2, Shift); |
69 | u = v_as_u64_f64 (z); |
70 | n = z - Shift; |
71 | |
72 | /* r = x - n*ln2/N. */ |
73 | r = x; |
74 | r = v_fma_f64 (-Ln2hi, n, r); |
75 | r = v_fma_f64 (-Ln2lo, n, r); |
76 | |
77 | e = u << (52 - V_EXP_TABLE_BITS); |
78 | i = u & IndexMask; |
79 | |
80 | /* y = exp(r) - 1 ~= r + C1 r^2 + C2 r^3 + C3 r^4. */ |
81 | r2 = r * r; |
82 | y = v_fma_f64 (C2, r, C1); |
83 | y = v_fma_f64 (C3, r2, y); |
84 | y = v_fma_f64 (y, r2, r); |
85 | |
86 | /* s = 2^(n/N). */ |
87 | u = v_lookup_u64 (Tab, i); |
88 | s = v_as_f64_u64 (u + e); |
89 | |
90 | if (unlikely (v_any_u64 (cmp))) |
91 | return specialcase (s, y, n); |
92 | return v_fma_f64 (y, s, s); |
93 | } |
94 | VPCS_ALIAS |
95 | #endif |
96 | |