1 | /* |
2 | * Single-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 | |
13 | static const float Poly[] = { |
14 | /* maxerr: 1.45358 +0.5 ulp. */ |
15 | 0x1.0e4020p-7f, |
16 | 0x1.573e2ep-5f, |
17 | 0x1.555e66p-3f, |
18 | 0x1.fffdb6p-2f, |
19 | 0x1.ffffecp-1f, |
20 | }; |
21 | #define C0 v_f32 (Poly[0]) |
22 | #define C1 v_f32 (Poly[1]) |
23 | #define C2 v_f32 (Poly[2]) |
24 | #define C3 v_f32 (Poly[3]) |
25 | #define C4 v_f32 (Poly[4]) |
26 | |
27 | #define Shift v_f32 (0x1.8p23f) |
28 | #define InvLn2 v_f32 (0x1.715476p+0f) |
29 | #define Ln2hi v_f32 (0x1.62e4p-1f) |
30 | #define Ln2lo v_f32 (0x1.7f7d1cp-20f) |
31 | |
32 | VPCS_ATTR |
33 | static v_f32_t |
34 | specialcase (v_f32_t poly, v_f32_t n, v_u32_t e, v_f32_t absn, v_u32_t cmp1, v_f32_t scale) |
35 | { |
36 | /* 2^n may overflow, break it up into s1*s2. */ |
37 | v_u32_t b = v_cond_u32 (x: n <= v_f32 (x: 0.0f)) & v_u32 (x: 0x82000000); |
38 | v_f32_t s1 = v_as_f32_u32 (x: v_u32 (x: 0x7f000000) + b); |
39 | v_f32_t s2 = v_as_f32_u32 (x: e - b); |
40 | v_u32_t cmp2 = v_cond_u32 (x: absn > v_f32 (x: 192.0f)); |
41 | v_u32_t r2 = v_as_u32_f32 (x: s1 * s1); |
42 | v_u32_t r1 = v_as_u32_f32 (x: v_fma_f32 (x: poly, y: s2, z: s2) * s1); |
43 | /* Similar to r1 but avoids double rounding in the subnormal range. */ |
44 | v_u32_t r0 = v_as_u32_f32 (x: v_fma_f32 (x: poly, y: scale, z: scale)); |
45 | return v_as_f32_u32 (x: (cmp2 & r2) | (~cmp2 & cmp1 & r1) | (~cmp1 & r0)); |
46 | } |
47 | |
48 | VPCS_ATTR |
49 | v_f32_t |
50 | V_NAME(expf) (v_f32_t x) |
51 | { |
52 | v_f32_t n, r, r2, scale, p, q, poly, absn, z; |
53 | v_u32_t cmp, e; |
54 | |
55 | /* exp(x) = 2^n (1 + poly(r)), with 1 + poly(r) in [1/sqrt(2),sqrt(2)] |
56 | x = ln2*n + r, with r in [-ln2/2, ln2/2]. */ |
57 | #if 1 |
58 | z = v_fma_f32 (x, InvLn2, Shift); |
59 | n = z - Shift; |
60 | r = v_fma_f32 (x: n, y: -Ln2hi, z: x); |
61 | r = v_fma_f32 (x: n, y: -Ln2lo, z: r); |
62 | e = v_as_u32_f32 (x: z) << 23; |
63 | #else |
64 | z = x * InvLn2; |
65 | n = v_round_f32 (z); |
66 | r = v_fma_f32 (n, -Ln2hi, x); |
67 | r = v_fma_f32 (n, -Ln2lo, r); |
68 | e = v_as_u32_s32 (v_round_s32 (z)) << 23; |
69 | #endif |
70 | scale = v_as_f32_u32 (x: e + v_u32 (x: 0x3f800000)); |
71 | absn = v_abs_f32 (x: n); |
72 | cmp = v_cond_u32 (x: absn > v_f32 (x: 126.0f)); |
73 | r2 = r * r; |
74 | p = v_fma_f32 (C0, y: r, C1); |
75 | q = v_fma_f32 (C2, y: r, C3); |
76 | q = v_fma_f32 (x: p, y: r2, z: q); |
77 | p = C4 * r; |
78 | poly = v_fma_f32 (x: q, y: r2, z: p); |
79 | if (unlikely (v_any_u32 (cmp))) |
80 | return specialcase (poly, n, e, absn, cmp1: cmp, scale); |
81 | return v_fma_f32 (x: poly, y: scale, z: scale); |
82 | } |
83 | VPCS_ALIAS |
84 | #endif |
85 | |