1 /*
2 "A Precision Approximation of the Gamma Function" - Cornelius Lanczos (1964)
3 "Lanczos Implementation of the Gamma Function" - Paul Godfrey (2001)
4 "An Analysis of the Lanczos Gamma Approximation" - Glendon Ralph Pugh (2004)
5 
6 approximation method:
7 
8                         (x - 0.5)         S(x)
9 Gamma(x) = (x + g - 0.5)         *  ----------------
10                                     exp(x + g - 0.5)
11 
12 with
13                  a1      a2      a3            aN
14 S(x) ~= [ a0 + ----- + ----- + ----- + ... + ----- ]
15                x + 1   x + 2   x + 3         x + N
16 
17 with a0, a1, a2, a3,.. aN constants which depend on g.
18 
19 for x < 0 the following reflection formula is used:
20 
21 Gamma(x)*Gamma(-x) = -pi/(x sin(pi x))
22 
23 most ideas and constants are from boost and python
24 */
25 #include "libm.h"
26 
27 static const double pi = 3.141592653589793238462643383279502884;
28 
29 /* sin(pi x) with x > 0x1p-100, if sin(pi*x)==0 the sign is arbitrary */
sinpi(double x)30 static double sinpi(double x)
31 {
32 	int n;
33 
34 	/* argument reduction: x = |x| mod 2 */
35 	/* spurious inexact when x is odd int */
36 	x = x * 0.5;
37 	x = 2 * (x - floor(x));
38 
39 	/* reduce x into [-.25,.25] */
40 	n = 4 * x;
41 	n = (n+1)/2;
42 	x -= n * 0.5;
43 
44 	x *= pi;
45 	switch (n) {
46 	default: /* case 4 */
47 	case 0:
48 		return __sin(x, 0, 0);
49 	case 1:
50 		return __cos(x, 0);
51 	case 2:
52 		return __sin(-x, 0, 0);
53 	case 3:
54 		return -__cos(x, 0);
55 	}
56 }
57 
58 #define N 12
59 //static const double g = 6.024680040776729583740234375;
60 static const double gmhalf = 5.524680040776729583740234375;
61 static const double Snum[N+1] = {
62 	23531376880.410759688572007674451636754734846804940,
63 	42919803642.649098768957899047001988850926355848959,
64 	35711959237.355668049440185451547166705960488635843,
65 	17921034426.037209699919755754458931112671403265390,
66 	6039542586.3520280050642916443072979210699388420708,
67 	1439720407.3117216736632230727949123939715485786772,
68 	248874557.86205415651146038641322942321632125127801,
69 	31426415.585400194380614231628318205362874684987640,
70 	2876370.6289353724412254090516208496135991145378768,
71 	186056.26539522349504029498971604569928220784236328,
72 	8071.6720023658162106380029022722506138218516325024,
73 	210.82427775157934587250973392071336271166969580291,
74 	2.5066282746310002701649081771338373386264310793408,
75 };
76 static const double Sden[N+1] = {
77 	0, 39916800, 120543840, 150917976, 105258076, 45995730, 13339535,
78 	2637558, 357423, 32670, 1925, 66, 1,
79 };
80 /* n! for small integer n */
81 static const double fact[] = {
82 	1, 1, 2, 6, 24, 120, 720, 5040.0, 40320.0, 362880.0, 3628800.0, 39916800.0,
83 	479001600.0, 6227020800.0, 87178291200.0, 1307674368000.0, 20922789888000.0,
84 	355687428096000.0, 6402373705728000.0, 121645100408832000.0,
85 	2432902008176640000.0, 51090942171709440000.0, 1124000727777607680000.0,
86 };
87 
88 /* S(x) rational function for positive x */
S(double x)89 static double S(double x)
90 {
91 	double_t num = 0, den = 0;
92 	int i;
93 
94 	/* to avoid overflow handle large x differently */
95 	if (x < 8)
96 		for (i = N; i >= 0; i--) {
97 			num = num * x + Snum[i];
98 			den = den * x + Sden[i];
99 		}
100 	else
101 		for (i = 0; i <= N; i++) {
102 			num = num / x + Snum[i];
103 			den = den / x + Sden[i];
104 		}
105 	return num/den;
106 }
107 
tgamma(double x)108 double tgamma(double x)
109 {
110 	union {double f; uint64_t i;} u = {x};
111 	double absx, y;
112 	double_t dy, z, r;
113 	uint32_t ix = u.i>>32 & 0x7fffffff;
114 	int sign = u.i>>63;
115 
116 	/* special cases */
117 	if (ix >= 0x7ff00000)
118 		/* tgamma(nan)=nan, tgamma(inf)=inf, tgamma(-inf)=nan with invalid */
119 		return x + INFINITY;
120 	if (ix < (0x3ff-54)<<20)
121 		/* |x| < 2^-54: tgamma(x) ~ 1/x, +-0 raises div-by-zero */
122 		return 1/x;
123 
124 	/* integer arguments */
125 	/* raise inexact when non-integer */
126 	if (x == floor(x)) {
127 		if (sign)
128 			return 0/0.0;
129 		if (x <= sizeof fact/sizeof *fact)
130 			return fact[(int)x - 1];
131 	}
132 
133 	/* x >= 172: tgamma(x)=inf with overflow */
134 	/* x =< -184: tgamma(x)=+-0 with underflow */
135 	if (ix >= 0x40670000) { /* |x| >= 184 */
136 		if (sign) {
137 			FORCE_EVAL((float)(0x1p-126/x));
138 			if (floor(x) * 0.5 == floor(x * 0.5))
139 				return 0;
140 			return -0.0;
141 		}
142 		x *= 0x1p1023;
143 		return x;
144 	}
145 
146 	absx = sign ? -x : x;
147 
148 	/* handle the error of x + g - 0.5 */
149 	y = absx + gmhalf;
150 	if (absx > gmhalf) {
151 		dy = y - absx;
152 		dy -= gmhalf;
153 	} else {
154 		dy = y - gmhalf;
155 		dy -= absx;
156 	}
157 
158 	z = absx - 0.5;
159 	r = S(absx) * exp(-y);
160 	if (x < 0) {
161 		/* reflection formula for negative x */
162 		/* sinpi(absx) is not 0, integers are already handled */
163 		r = -pi / (sinpi(absx) * absx * r);
164 		dy = -dy;
165 		z = -z;
166 	}
167 	r += dy * (gmhalf+0.5) * r / y;
168 	z = pow(y, 0.5*z);
169 	y = r * z * z;
170 	return y;
171 }
172 
173 #if 0
174 double __lgamma_r(double x, int *sign)
175 {
176 	double r, absx;
177 
178 	*sign = 1;
179 
180 	/* special cases */
181 	if (!isfinite(x))
182 		/* lgamma(nan)=nan, lgamma(+-inf)=inf */
183 		return x*x;
184 
185 	/* integer arguments */
186 	if (x == floor(x) && x <= 2) {
187 		/* n <= 0: lgamma(n)=inf with divbyzero */
188 		/* n == 1,2: lgamma(n)=0 */
189 		if (x <= 0)
190 			return 1/0.0;
191 		return 0;
192 	}
193 
194 	absx = fabs(x);
195 
196 	/* lgamma(x) ~ -log(|x|) for tiny |x| */
197 	if (absx < 0x1p-54) {
198 		*sign = 1 - 2*!!signbit(x);
199 		return -log(absx);
200 	}
201 
202 	/* use tgamma for smaller |x| */
203 	if (absx < 128) {
204 		x = tgamma(x);
205 		*sign = 1 - 2*!!signbit(x);
206 		return log(fabs(x));
207 	}
208 
209 	/* second term (log(S)-g) could be more precise here.. */
210 	/* or with stirling: (|x|-0.5)*(log(|x|)-1) + poly(1/|x|) */
211 	r = (absx-0.5)*(log(absx+gmhalf)-1) + (log(S(absx)) - (gmhalf+0.5));
212 	if (x < 0) {
213 		/* reflection formula for negative x */
214 		x = sinpi(absx);
215 		*sign = 2*!!signbit(x) - 1;
216 		r = log(pi/(fabs(x)*absx)) - r;
217 	}
218 	return r;
219 }
220 
221 weak_alias(__lgamma_r, lgamma_r);
222 #endif
223