xref: /freebsd/lib/msun/src/k_log.h (revision a91a246563dffa876a52f53a98de4af9fa364c52)
1 
2 /*
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  *
6  * Developed at SunSoft, a Sun Microsystems, Inc. business.
7  * Permission to use, copy, modify, and distribute this
8  * software is freely granted, provided that this notice
9  * is preserved.
10  * ====================================================
11  */
12 
13 /*
14  * k_log1p(f):
15  * Return log(1+f) - f for 1+f in ~[sqrt(2)/2, sqrt(2)].
16  *
17  * The following describes the overall strategy for computing
18  * logarithms in base e.  The argument reduction and adding the final
19  * term of the polynomial are done by the caller for increased accuracy
20  * when different bases are used.
21  *
22  * Method :
23  *   1. Argument Reduction: find k and f such that
24  *			x = 2^k * (1+f),
25  *	   where  sqrt(2)/2 < 1+f < sqrt(2) .
26  *
27  *   2. Approximation of log(1+f).
28  *	Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
29  *		 = 2s + 2/3 s**3 + 2/5 s**5 + .....,
30  *	     	 = 2s + s*R
31  *      We use a special Reme algorithm on [0,0.1716] to generate
32  * 	a polynomial of degree 14 to approximate R The maximum error
33  *	of this polynomial approximation is bounded by 2**-58.45. In
34  *	other words,
35  *		        2      4      6      8      10      12      14
36  *	    R(z) ~ Lg1*s +Lg2*s +Lg3*s +Lg4*s +Lg5*s  +Lg6*s  +Lg7*s
37  *  	(the values of Lg1 to Lg7 are listed in the program)
38  *	and
39  *	    |      2          14          |     -58.45
40  *	    | Lg1*s +...+Lg7*s    -  R(z) | <= 2
41  *	    |                             |
42  *	Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2.
43  *	In order to guarantee error in log below 1ulp, we compute log
44  *	by
45  *		log(1+f) = f - s*(f - R)	(if f is not too large)
46  *		log(1+f) = f - (hfsq - s*(hfsq+R)).	(better accuracy)
47  *
48  *	3. Finally,  log(x) = k*ln2 + log(1+f).
49  *			    = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo)))
50  *	   Here ln2 is split into two floating point number:
51  *			ln2_hi + ln2_lo,
52  *	   where n*ln2_hi is always exact for |n| < 2000.
53  *
54  * Special cases:
55  *	log(x) is NaN with signal if x < 0 (including -INF) ;
56  *	log(+INF) is +INF; log(0) is -INF with signal;
57  *	log(NaN) is that NaN with no signal.
58  *
59  * Accuracy:
60  *	according to an error analysis, the error is always less than
61  *	1 ulp (unit in the last place).
62  *
63  * Constants:
64  * The hexadecimal values are the intended ones for the following
65  * constants. The decimal values may be used, provided that the
66  * compiler will convert from decimal to binary accurately enough
67  * to produce the hexadecimal values shown.
68  */
69 
70 static const double
71 Lg1 = 6.666666666666735130e-01,  /* 3FE55555 55555593 */
72 Lg2 = 3.999999999940941908e-01,  /* 3FD99999 9997FA04 */
73 Lg3 = 2.857142874366239149e-01,  /* 3FD24924 94229359 */
74 Lg4 = 2.222219843214978396e-01,  /* 3FCC71C5 1D8E78AF */
75 Lg5 = 1.818357216161805012e-01,  /* 3FC74664 96CB03DE */
76 Lg6 = 1.531383769920937332e-01,  /* 3FC39A09 D078C69F */
77 Lg7 = 1.479819860511658591e-01;  /* 3FC2F112 DF3E5244 */
78 
79 /*
80  * We always inline k_log1p(), since doing so produces a
81  * substantial performance improvement (~40% on amd64).
82  */
83 static inline double
84 k_log1p(double f)
85 {
86 	double hfsq,s,z,R,w,t1,t2;
87 
88  	s = f/(2.0+f);
89 	z = s*s;
90 	w = z*z;
91 	t1= w*(Lg2+w*(Lg4+w*Lg6));
92 	t2= z*(Lg1+w*(Lg3+w*(Lg5+w*Lg7)));
93 	R = t2+t1;
94 	hfsq=0.5*f*f;
95 	return s*(hfsq+R);
96 }
97