1 | /* |
2 | * Single-precision vector log 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 | /* 3.34 ulp error */ |
15 | -0x1.3e737cp-3f, 0x1.5a9aa2p-3f, -0x1.4f9934p-3f, 0x1.961348p-3f, |
16 | -0x1.00187cp-2f, 0x1.555d7cp-2f, -0x1.ffffc8p-2f, |
17 | }; |
18 | #define P7 v_f32 (Poly[0]) |
19 | #define P6 v_f32 (Poly[1]) |
20 | #define P5 v_f32 (Poly[2]) |
21 | #define P4 v_f32 (Poly[3]) |
22 | #define P3 v_f32 (Poly[4]) |
23 | #define P2 v_f32 (Poly[5]) |
24 | #define P1 v_f32 (Poly[6]) |
25 | |
26 | #define Ln2 v_f32 (0x1.62e43p-1f) /* 0x3f317218 */ |
27 | #define Min v_u32 (0x00800000) |
28 | #define Max v_u32 (0x7f800000) |
29 | #define Mask v_u32 (0x007fffff) |
30 | #define Off v_u32 (0x3f2aaaab) /* 0.666667 */ |
31 | |
32 | VPCS_ATTR |
33 | __attribute__ ((noinline)) static v_f32_t |
34 | specialcase (v_f32_t x, v_f32_t y, v_u32_t cmp) |
35 | { |
36 | /* Fall back to scalar code. */ |
37 | return v_call_f32 (f: logf, x, y, p: cmp); |
38 | } |
39 | |
40 | VPCS_ATTR |
41 | v_f32_t |
42 | V_NAME(logf) (v_f32_t x) |
43 | { |
44 | v_f32_t n, p, q, r, r2, y; |
45 | v_u32_t u, cmp; |
46 | |
47 | u = v_as_u32_f32 (x); |
48 | cmp = v_cond_u32 (x: u - Min >= Max - Min); |
49 | |
50 | /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3 */ |
51 | u -= Off; |
52 | n = v_to_f32_s32 (x: v_as_s32_u32 (x: u) >> 23); /* signextend */ |
53 | u &= Mask; |
54 | u += Off; |
55 | r = v_as_f32_u32 (x: u) - v_f32 (x: 1.0f); |
56 | |
57 | /* y = log(1+r) + n*ln2. */ |
58 | r2 = r * r; |
59 | /* n*ln2 + r + r2*(P1 + r*P2 + r2*(P3 + r*P4 + r2*(P5 + r*P6 + r2*P7))). */ |
60 | p = v_fma_f32 (P6, y: r, P5); |
61 | q = v_fma_f32 (P4, y: r, P3); |
62 | y = v_fma_f32 (P2, y: r, P1); |
63 | p = v_fma_f32 (P7, y: r2, z: p); |
64 | q = v_fma_f32 (x: p, y: r2, z: q); |
65 | y = v_fma_f32 (x: q, y: r2, z: y); |
66 | p = v_fma_f32 (Ln2, y: n, z: r); |
67 | y = v_fma_f32 (x: y, y: r2, z: p); |
68 | |
69 | if (unlikely (v_any_u32 (cmp))) |
70 | return specialcase (x, y, cmp); |
71 | return y; |
72 | } |
73 | VPCS_ALIAS |
74 | #endif |
75 | |