1 | /* |
2 | * ==================================================== |
3 | * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
4 | * |
5 | * Developed at SunPro, a Sun Microsystems, Inc. business. |
6 | * Permission to use, copy, modify, and distribute this |
7 | * software is freely granted, provided that this notice |
8 | * is preserved. |
9 | * ==================================================== |
10 | */ |
11 | |
12 | /* __ieee754_log2(x) |
13 | * Return the logarithm to base 2 of x |
14 | * |
15 | * Method : |
16 | * 1. Argument Reduction: find k and f such that |
17 | * x = 2^k * (1+f), |
18 | * where sqrt(2)/2 < 1+f < sqrt(2) . |
19 | * |
20 | * 2. Approximation of log(1+f). |
21 | * Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s) |
22 | * = 2s + 2/3 s**3 + 2/5 s**5 + ....., |
23 | * = 2s + s*R |
24 | * We use a special Reme algorithm on [0,0.1716] to generate |
25 | * a polynomial of degree 14 to approximate R The maximum error |
26 | * of this polynomial approximation is bounded by 2**-58.45. In |
27 | * other words, |
28 | * 2 4 6 8 10 12 14 |
29 | * R(z) ~ Lg1*s +Lg2*s +Lg3*s +Lg4*s +Lg5*s +Lg6*s +Lg7*s |
30 | * (the values of Lg1 to Lg7 are listed in the program) |
31 | * and |
32 | * | 2 14 | -58.45 |
33 | * | Lg1*s +...+Lg7*s - R(z) | <= 2 |
34 | * | | |
35 | * Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2. |
36 | * In order to guarantee error in log below 1ulp, we compute log |
37 | * by |
38 | * log(1+f) = f - s*(f - R) (if f is not too large) |
39 | * log(1+f) = f - (hfsq - s*(hfsq+R)). (better accuracy) |
40 | * |
41 | * 3. Finally, log(x) = k + log(1+f). |
42 | * = k+(f-(hfsq-(s*(hfsq+R)))) |
43 | * |
44 | * Special cases: |
45 | * log2(x) is NaN with signal if x < 0 (including -INF) ; |
46 | * log2(+INF) is +INF; log(0) is -INF with signal; |
47 | * log2(NaN) is that NaN with no signal. |
48 | * |
49 | * Constants: |
50 | * The hexadecimal values are the intended ones for the following |
51 | * constants. The decimal values may be used, provided that the |
52 | * compiler will convert from decimal to binary accurately enough |
53 | * to produce the hexadecimal values shown. |
54 | */ |
55 | |
56 | #include <math.h> |
57 | #include <math_private.h> |
58 | |
59 | static const double ln2 = 0.69314718055994530942; |
60 | static const double two54 = 1.80143985094819840000e+16; /* 4350000000000000 */ |
61 | static const double Lg1 = 6.666666666666735130e-01; /* 3FE5555555555593 */ |
62 | static const double Lg2 = 3.999999999940941908e-01; /* 3FD999999997FA04 */ |
63 | static const double Lg3 = 2.857142874366239149e-01; /* 3FD2492494229359 */ |
64 | static const double Lg4 = 2.222219843214978396e-01; /* 3FCC71C51D8E78AF */ |
65 | static const double Lg5 = 1.818357216161805012e-01; /* 3FC7466496CB03DE */ |
66 | static const double Lg6 = 1.531383769920937332e-01; /* 3FC39A09D078C69F */ |
67 | static const double Lg7 = 1.479819860511658591e-01; /* 3FC2F112DF3E5244 */ |
68 | |
69 | static const double zero = 0.0; |
70 | |
71 | double |
72 | __ieee754_log2 (double x) |
73 | { |
74 | double hfsq, f, s, z, R, w, t1, t2, dk; |
75 | int64_t hx, i, j; |
76 | int32_t k; |
77 | |
78 | EXTRACT_WORDS64 (hx, x); |
79 | |
80 | k = 0; |
81 | if (hx < INT64_C(0x0010000000000000)) |
82 | { /* x < 2**-1022 */ |
83 | if (__glibc_unlikely ((hx & UINT64_C(0x7fffffffffffffff)) == 0)) |
84 | return -two54 / (x - x); /* log(+-0)=-inf */ |
85 | if (__glibc_unlikely (hx < 0)) |
86 | return (x - x) / (x - x); /* log(-#) = NaN */ |
87 | k -= 54; |
88 | x *= two54; /* subnormal number, scale up x */ |
89 | EXTRACT_WORDS64 (hx, x); |
90 | } |
91 | if (__glibc_unlikely (hx >= UINT64_C(0x7ff0000000000000))) |
92 | return x + x; |
93 | k += (hx >> 52) - 1023; |
94 | hx &= UINT64_C(0x000fffffffffffff); |
95 | i = (hx + UINT64_C(0x95f6400000000)) & UINT64_C(0x10000000000000); |
96 | /* normalize x or x/2 */ |
97 | INSERT_WORDS64 (x, hx | (i ^ UINT64_C(0x3ff0000000000000))); |
98 | k += (i >> 52); |
99 | dk = (double) k; |
100 | f = x - 1.0; |
101 | if ((UINT64_C(0x000fffffffffffff) & (2 + hx)) < 3) |
102 | { /* |f| < 2**-20 */ |
103 | if (f == zero) |
104 | return dk; |
105 | R = f * f * (0.5 - 0.33333333333333333 * f); |
106 | return dk - (R - f) / ln2; |
107 | } |
108 | s = f / (2.0 + f); |
109 | z = s * s; |
110 | i = hx - UINT64_C(0x6147a00000000); |
111 | w = z * z; |
112 | j = UINT64_C(0x6b85100000000) - hx; |
113 | t1 = w * (Lg2 + w * (Lg4 + w * Lg6)); |
114 | t2 = z * (Lg1 + w * (Lg3 + w * (Lg5 + w * Lg7))); |
115 | i |= j; |
116 | R = t2 + t1; |
117 | if (i > 0) |
118 | { |
119 | hfsq = 0.5 * f * f; |
120 | return dk - ((hfsq - (s * (hfsq + R))) - f) / ln2; |
121 | } |
122 | else |
123 | { |
124 | return dk - ((s * (f - R)) - f) / ln2; |
125 | } |
126 | } |
127 | |
128 | strong_alias (__ieee754_log2, __log2_finite) |
129 | |