xref: /illumos-gate/usr/src/lib/libm/common/C/log1p.c (revision a6bde1a23b60f140c7ed78df979c2e22b1ed9b2c)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
23  */
24 /*
25  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
26  * Use is subject to license terms.
27  */
28 
29 #pragma weak log1p = __log1p
30 
31 /* INDENT OFF */
32 /*
33  * Method :
34  *   1. Argument Reduction: find k and f such that
35  *			1+x = 2^k * (1+f),
36  *	   where  sqrt(2)/2 < 1+f < sqrt(2) .
37  *
38  *      Note. If k=0, then f=x is exact. However, if k != 0, then f
39  *	may not be representable exactly. In that case, a correction
40  *	term is need. Let u=1+x rounded. Let c = (1+x)-u, then
41  *	log(1+x) - log(u) ~ c/u. Thus, we proceed to compute log(u),
42  *	and add back the correction term c/u.
43  *	(Note: when x > 2**53, one can simply return log(x))
44  *
45  *   2. Approximation of log1p(f).
46  *	Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
47  *		 = 2s + 2/3 s**3 + 2/5 s**5 + .....,
48  *		 = 2s + s*R
49  *      We use a special Reme algorithm on [0,0.1716] to generate
50  * 	a polynomial of degree 14 to approximate R The maximum error
51  *	of this polynomial approximation is bounded by 2**-58.45. In
52  *	other words,
53  *		        2      4      6      8      10      12      14
54  *	    R(z) ~ Lp1*s +Lp2*s +Lp3*s +Lp4*s +Lp5*s  +Lp6*s  +Lp7*s
55  *  	(the values of Lp1 to Lp7 are listed in the program)
56  *	and
57  *	    |      2          14          |     -58.45
58  *	    | Lp1*s +...+Lp7*s    -  R(z) | <= 2
59  *	    |                             |
60  *	Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2.
61  *	In order to guarantee error in log below 1ulp, we compute log
62  *	by
63  *		log1p(f) = f - (hfsq - s*(hfsq+R)).
64  *
65  *	3. Finally, log1p(x) = k*ln2 + log1p(f).
66  *			     = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo)))
67  *	   Here ln2 is splitted into two floating point number:
68  *			ln2_hi + ln2_lo,
69  *	   where n*ln2_hi is always exact for |n| < 2000.
70  *
71  * Special cases:
72  *	log1p(x) is NaN with signal if x < -1 (including -INF) ;
73  *	log1p(+INF) is +INF; log1p(-1) is -INF with signal;
74  *	log1p(NaN) is that NaN with no signal.
75  *
76  * Accuracy:
77  *	according to an error analysis, the error is always less than
78  *	1 ulp (unit in the last place).
79  *
80  * Constants:
81  * The hexadecimal values are the intended ones for the following
82  * constants. The decimal values may be used, provided that the
83  * compiler will convert from decimal to binary accurately enough
84  * to produce the hexadecimal values shown.
85  *
86  * Note: Assuming log() return accurate answer, the following
87  *	 algorithm can be used to compute log1p(x) to within a few ULP:
88  *
89  *		u = 1+x;
90  *		if (u == 1.0) return x ; else
91  *			   return log(u)*(x/(u-1.0));
92  *
93  *	 See HP-15C Advanced Functions Handbook, p.193.
94  */
95 /* INDENT ON */
96 
97 #include "libm.h"
98 
99 static const double xxx[] = {
100 /* ln2_hi */	6.93147180369123816490e-01,	/* 3fe62e42 fee00000 */
101 /* ln2_lo */	1.90821492927058770002e-10,	/* 3dea39ef 35793c76 */
102 /* two54 */	1.80143985094819840000e+16,	/* 43500000 00000000 */
103 /* Lp1 */	6.666666666666735130e-01,	/* 3FE55555 55555593 */
104 /* Lp2 */	3.999999999940941908e-01,	/* 3FD99999 9997FA04 */
105 /* Lp3 */	2.857142874366239149e-01,	/* 3FD24924 94229359 */
106 /* Lp4 */	2.222219843214978396e-01,	/* 3FCC71C5 1D8E78AF */
107 /* Lp5 */	1.818357216161805012e-01,	/* 3FC74664 96CB03DE */
108 /* Lp6 */	1.531383769920937332e-01,	/* 3FC39A09 D078C69F */
109 /* Lp7 */	1.479819860511658591e-01,	/* 3FC2F112 DF3E5244 */
110 /* zero */	0.0
111 };
112 #define	ln2_hi	xxx[0]
113 #define	ln2_lo	xxx[1]
114 #define	two54	xxx[2]
115 #define	Lp1	xxx[3]
116 #define	Lp2	xxx[4]
117 #define	Lp3	xxx[5]
118 #define	Lp4	xxx[6]
119 #define	Lp5	xxx[7]
120 #define	Lp6	xxx[8]
121 #define	Lp7	xxx[9]
122 #define	zero	xxx[10]
123 
124 double
125 log1p(double x) {
126 	double	hfsq, f, c = 0.0, s, z, R, u;
127 	int	k, hx, hu, ax;
128 
129 	hx = ((int *)&x)[HIWORD];		/* high word of x */
130 	ax = hx & 0x7fffffff;
131 
132 	if (ax >= 0x7ff00000) { /* x is inf or nan */
133 		if (((hx - 0xfff00000) | ((int *)&x)[LOWORD]) == 0) /* -inf */
134 			return (_SVID_libm_err(x, x, 44));
135 		return (x * x);
136 	}
137 
138 	k = 1;
139 	if (hx < 0x3FDA827A) {	/* x < 0.41422  */
140 		if (ax >= 0x3ff00000)	/* x <= -1.0 */
141 			return (_SVID_libm_err(x, x, x == -1.0 ? 43 : 44));
142 		if (ax < 0x3e200000) {	/* |x| < 2**-29 */
143 			if (two54 + x > zero &&	/* raise inexact */
144 			    ax < 0x3c900000)	/* |x| < 2**-54 */
145 				return (x);
146 			else
147 				return (x - x * x * 0.5);
148 		}
149 		if (hx > 0 || hx <= (int)0xbfd2bec3) {	/* -0.2929<x<0.41422 */
150 			k = 0;
151 			f = x;
152 			hu = 1;
153 		}
154 	}
155 	/* We will initialize 'c' here. */
156 	if (k != 0) {
157 		if (hx < 0x43400000) {
158 			u = 1.0 + x;
159 			hu = ((int *)&u)[HIWORD];	/* high word of u */
160 			k = (hu >> 20) - 1023;
161 			/*
162 			 * correction term
163 			 */
164 			c = k > 0 ? 1.0 - (u - x) : x - (u - 1.0);
165 			c /= u;
166 		} else {
167 			u = x;
168 			hu = ((int *)&u)[HIWORD];	/* high word of u */
169 			k = (hu >> 20) - 1023;
170 			c = 0;
171 		}
172 		hu &= 0x000fffff;
173 		if (hu < 0x6a09e) {	/* normalize u */
174 			((int *)&u)[HIWORD] = hu | 0x3ff00000;
175 		} else {			/* normalize u/2 */
176 			k += 1;
177 			((int *)&u)[HIWORD] = hu | 0x3fe00000;
178 			hu = (0x00100000 - hu) >> 2;
179 		}
180 		f = u - 1.0;
181 	}
182 	hfsq = 0.5 * f * f;
183 	if (hu == 0) {		/* |f| < 2**-20 */
184 		if (f == zero) {
185 			if (k == 0)
186 				return (zero);
187 			/* We already initialized 'c' before, when (k != 0) */
188 			c += k * ln2_lo;
189 			return (k * ln2_hi + c);
190 		}
191 		R = hfsq * (1.0 - 0.66666666666666666 * f);
192 		if (k == 0)
193 			return (f - R);
194 		return (k * ln2_hi - ((R - (k * ln2_lo + c)) - f));
195 	}
196 	s = f / (2.0 + f);
197 	z = s * s;
198 	R = z * (Lp1 + z * (Lp2 + z * (Lp3 + z * (Lp4 + z * (Lp5 +
199 		z * (Lp6 + z * Lp7))))));
200 	if (k == 0)
201 		return (f - (hfsq - s * (hfsq + R)));
202 	return (k * ln2_hi - ((hfsq - (s * (hfsq + R) +
203 		(k * ln2_lo + c))) - f));
204 }
205