| 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 */ |
| 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 */ |
| 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 | |
| 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 |