1 | /* |
2 | * Single-precision log2 function. |
3 | * |
4 | * Copyright (c) 2017-2018, Arm Limited. |
5 | * SPDX-License-Identifier: MIT |
6 | */ |
7 | |
8 | #include <math.h> |
9 | #include <stdint.h> |
10 | #include "libm.h" |
11 | #include "log2f_data.h" |
12 | |
13 | /* |
14 | LOG2F_TABLE_BITS = 4 |
15 | LOG2F_POLY_ORDER = 4 |
16 | |
17 | ULP error: 0.752 (nearest rounding.) |
18 | Relative error: 1.9 * 2^-26 (before rounding.) |
19 | */ |
20 | |
21 | #define N (1 << LOG2F_TABLE_BITS) |
22 | #define T __log2f_data.tab |
23 | #define A __log2f_data.poly |
24 | #define OFF 0x3f330000 |
25 | |
26 | float log2f(float x) |
27 | { |
28 | double_t z, r, r2, p, y, y0, invc, logc; |
29 | uint32_t ix, iz, top, tmp; |
30 | int k, i; |
31 | |
32 | ix = asuint(x); |
33 | /* Fix sign of zero with downward rounding when x==1. */ |
34 | if (WANT_ROUNDING && predict_false(ix == 0x3f800000)) |
35 | return 0; |
36 | if (predict_false(ix - 0x00800000 >= 0x7f800000 - 0x00800000)) { |
37 | /* x < 0x1p-126 or inf or nan. */ |
38 | if (ix * 2 == 0) |
39 | return __math_divzerof(1); |
40 | if (ix == 0x7f800000) /* log2(inf) == inf. */ |
41 | return x; |
42 | if ((ix & 0x80000000) || ix * 2 >= 0xff000000) |
43 | return __math_invalidf(x); |
44 | /* x is subnormal, normalize it. */ |
45 | ix = asuint(x * 0x1p23f); |
46 | ix -= 23 << 23; |
47 | } |
48 | |
49 | /* x = 2^k z; where z is in range [OFF,2*OFF] and exact. |
50 | The range is split into N subintervals. |
51 | The ith subinterval contains z and c is near its center. */ |
52 | tmp = ix - OFF; |
53 | i = (tmp >> (23 - LOG2F_TABLE_BITS)) % N; |
54 | top = tmp & 0xff800000; |
55 | iz = ix - top; |
56 | k = (int32_t)tmp >> 23; /* arithmetic shift */ |
57 | invc = T[i].invc; |
58 | logc = T[i].logc; |
59 | z = (double_t)asfloat(iz); |
60 | |
61 | /* log2(x) = log1p(z/c-1)/ln2 + log2(c) + k */ |
62 | r = z * invc - 1; |
63 | y0 = logc + (double_t)k; |
64 | |
65 | /* Pipelined polynomial evaluation to approximate log1p(r)/ln2. */ |
66 | r2 = r * r; |
67 | y = A[1] * r + A[2]; |
68 | y = A[0] * r2 + y; |
69 | p = A[3] * r + y0; |
70 | y = y * r2 + p; |
71 | return eval_as_float(y); |
72 | } |
73 | |