xref: /freebsd/lib/msun/src/e_lgamma_r.c (revision 1d386b48a555f61cb7325543adbbb5c3f3407a66)
1 /* @(#)e_lgamma_r.c 1.3 95/01/18 */
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 #include <sys/cdefs.h>
14 /* lgamma_r(x, signgamp)
15  * Reentrant version of the logarithm of the Gamma function
16  * with user provide pointer for the sign of Gamma(x).
17  *
18  * Method:
19  *   1. Argument Reduction for 0 < x <= 8
20  * 	Since gamma(1+s)=s*gamma(s), for x in [0,8], we may
21  * 	reduce x to a number in [1.5,2.5] by
22  * 		lgamma(1+s) = log(s) + lgamma(s)
23  *	for example,
24  *		lgamma(7.3) = log(6.3) + lgamma(6.3)
25  *			    = log(6.3*5.3) + lgamma(5.3)
26  *			    = log(6.3*5.3*4.3*3.3*2.3) + lgamma(2.3)
27  *   2. Polynomial approximation of lgamma around its
28  *	minimum ymin=1.461632144968362245 to maintain monotonicity.
29  *	On [ymin-0.23, ymin+0.27] (i.e., [1.23164,1.73163]), use
30  *		Let z = x-ymin;
31  *		lgamma(x) = -1.214862905358496078218 + z^2*poly(z)
32  *	where
33  *		poly(z) is a 14 degree polynomial.
34  *   2. Rational approximation in the primary interval [2,3]
35  *	We use the following approximation:
36  *		s = x-2.0;
37  *		lgamma(x) = 0.5*s + s*P(s)/Q(s)
38  *	with accuracy
39  *		|P/Q - (lgamma(x)-0.5s)| < 2**-61.71
40  *	Our algorithms are based on the following observation
41  *
42  *                             zeta(2)-1    2    zeta(3)-1    3
43  * lgamma(2+s) = s*(1-Euler) + --------- * s  -  --------- * s  + ...
44  *                                 2                 3
45  *
46  *	where Euler = 0.5771... is the Euler constant, which is very
47  *	close to 0.5.
48  *
49  *   3. For x>=8, we have
50  *	lgamma(x)~(x-0.5)log(x)-x+0.5*log(2pi)+1/(12x)-1/(360x**3)+....
51  *	(better formula:
52  *	   lgamma(x)~(x-0.5)*(log(x)-1)-.5*(log(2pi)-1) + ...)
53  *	Let z = 1/x, then we approximation
54  *		f(z) = lgamma(x) - (x-0.5)(log(x)-1)
55  *	by
56  *	  			    3       5             11
57  *		w = w0 + w1*z + w2*z  + w3*z  + ... + w6*z
58  *	where
59  *		|w - f(z)| < 2**-58.74
60  *
61  *   4. For negative x, since (G is gamma function)
62  *		-x*G(-x)*G(x) = pi/sin(pi*x),
63  * 	we have
64  * 		G(x) = pi/(sin(pi*x)*(-x)*G(-x))
65  *	since G(-x) is positive, sign(G(x)) = sign(sin(pi*x)) for x<0
66  *	Hence, for x<0, signgam = sign(sin(pi*x)) and
67  *		lgamma(x) = log(|Gamma(x)|)
68  *			  = log(pi/(|x*sin(pi*x)|)) - lgamma(-x);
69  *	Note: one should avoid compute pi*(-x) directly in the
70  *	      computation of sin(pi*(-x)).
71  *
72  *   5. Special Cases
73  *		lgamma(2+s) ~ s*(1-Euler) for tiny s
74  *		lgamma(1) = lgamma(2) = 0
75  *		lgamma(x) ~ -log(|x|) for tiny x
76  *		lgamma(0) = lgamma(neg.integer) = inf and raise divide-by-zero
77  *		lgamma(inf) = inf
78  *		lgamma(-inf) = inf (bug for bug compatible with C99!?)
79  */
80 
81 #include <float.h>
82 
83 #include "math.h"
84 #include "math_private.h"
85 
86 static const volatile double vzero = 0;
87 
88 static const double
89 zero=  0.00000000000000000000e+00,
90 half=  5.00000000000000000000e-01, /* 0x3FE00000, 0x00000000 */
91 one =  1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */
92 pi  =  3.14159265358979311600e+00, /* 0x400921FB, 0x54442D18 */
93 a0  =  7.72156649015328655494e-02, /* 0x3FB3C467, 0xE37DB0C8 */
94 a1  =  3.22467033424113591611e-01, /* 0x3FD4A34C, 0xC4A60FAD */
95 a2  =  6.73523010531292681824e-02, /* 0x3FB13E00, 0x1A5562A7 */
96 a3  =  2.05808084325167332806e-02, /* 0x3F951322, 0xAC92547B */
97 a4  =  7.38555086081402883957e-03, /* 0x3F7E404F, 0xB68FEFE8 */
98 a5  =  2.89051383673415629091e-03, /* 0x3F67ADD8, 0xCCB7926B */
99 a6  =  1.19270763183362067845e-03, /* 0x3F538A94, 0x116F3F5D */
100 a7  =  5.10069792153511336608e-04, /* 0x3F40B6C6, 0x89B99C00 */
101 a8  =  2.20862790713908385557e-04, /* 0x3F2CF2EC, 0xED10E54D */
102 a9  =  1.08011567247583939954e-04, /* 0x3F1C5088, 0x987DFB07 */
103 a10 =  2.52144565451257326939e-05, /* 0x3EFA7074, 0x428CFA52 */
104 a11 =  4.48640949618915160150e-05, /* 0x3F07858E, 0x90A45837 */
105 tc  =  1.46163214496836224576e+00, /* 0x3FF762D8, 0x6356BE3F */
106 tf  = -1.21486290535849611461e-01, /* 0xBFBF19B9, 0xBCC38A42 */
107 /* tt = -(tail of tf) */
108 tt  = -3.63867699703950536541e-18, /* 0xBC50C7CA, 0xA48A971F */
109 t0  =  4.83836122723810047042e-01, /* 0x3FDEF72B, 0xC8EE38A2 */
110 t1  = -1.47587722994593911752e-01, /* 0xBFC2E427, 0x8DC6C509 */
111 t2  =  6.46249402391333854778e-02, /* 0x3FB08B42, 0x94D5419B */
112 t3  = -3.27885410759859649565e-02, /* 0xBFA0C9A8, 0xDF35B713 */
113 t4  =  1.79706750811820387126e-02, /* 0x3F9266E7, 0x970AF9EC */
114 t5  = -1.03142241298341437450e-02, /* 0xBF851F9F, 0xBA91EC6A */
115 t6  =  6.10053870246291332635e-03, /* 0x3F78FCE0, 0xE370E344 */
116 t7  = -3.68452016781138256760e-03, /* 0xBF6E2EFF, 0xB3E914D7 */
117 t8  =  2.25964780900612472250e-03, /* 0x3F6282D3, 0x2E15C915 */
118 t9  = -1.40346469989232843813e-03, /* 0xBF56FE8E, 0xBF2D1AF1 */
119 t10 =  8.81081882437654011382e-04, /* 0x3F4CDF0C, 0xEF61A8E9 */
120 t11 = -5.38595305356740546715e-04, /* 0xBF41A610, 0x9C73E0EC */
121 t12 =  3.15632070903625950361e-04, /* 0x3F34AF6D, 0x6C0EBBF7 */
122 t13 = -3.12754168375120860518e-04, /* 0xBF347F24, 0xECC38C38 */
123 t14 =  3.35529192635519073543e-04, /* 0x3F35FD3E, 0xE8C2D3F4 */
124 u0  = -7.72156649015328655494e-02, /* 0xBFB3C467, 0xE37DB0C8 */
125 u1  =  6.32827064025093366517e-01, /* 0x3FE4401E, 0x8B005DFF */
126 u2  =  1.45492250137234768737e+00, /* 0x3FF7475C, 0xD119BD6F */
127 u3  =  9.77717527963372745603e-01, /* 0x3FEF4976, 0x44EA8450 */
128 u4  =  2.28963728064692451092e-01, /* 0x3FCD4EAE, 0xF6010924 */
129 u5  =  1.33810918536787660377e-02, /* 0x3F8B678B, 0xBF2BAB09 */
130 v1  =  2.45597793713041134822e+00, /* 0x4003A5D7, 0xC2BD619C */
131 v2  =  2.12848976379893395361e+00, /* 0x40010725, 0xA42B18F5 */
132 v3  =  7.69285150456672783825e-01, /* 0x3FE89DFB, 0xE45050AF */
133 v4  =  1.04222645593369134254e-01, /* 0x3FBAAE55, 0xD6537C88 */
134 v5  =  3.21709242282423911810e-03, /* 0x3F6A5ABB, 0x57D0CF61 */
135 s0  = -7.72156649015328655494e-02, /* 0xBFB3C467, 0xE37DB0C8 */
136 s1  =  2.14982415960608852501e-01, /* 0x3FCB848B, 0x36E20878 */
137 s2  =  3.25778796408930981787e-01, /* 0x3FD4D98F, 0x4F139F59 */
138 s3  =  1.46350472652464452805e-01, /* 0x3FC2BB9C, 0xBEE5F2F7 */
139 s4  =  2.66422703033638609560e-02, /* 0x3F9B481C, 0x7E939961 */
140 s5  =  1.84028451407337715652e-03, /* 0x3F5E26B6, 0x7368F239 */
141 s6  =  3.19475326584100867617e-05, /* 0x3F00BFEC, 0xDD17E945 */
142 r1  =  1.39200533467621045958e+00, /* 0x3FF645A7, 0x62C4AB74 */
143 r2  =  7.21935547567138069525e-01, /* 0x3FE71A18, 0x93D3DCDC */
144 r3  =  1.71933865632803078993e-01, /* 0x3FC601ED, 0xCCFBDF27 */
145 r4  =  1.86459191715652901344e-02, /* 0x3F9317EA, 0x742ED475 */
146 r5  =  7.77942496381893596434e-04, /* 0x3F497DDA, 0xCA41A95B */
147 r6  =  7.32668430744625636189e-06, /* 0x3EDEBAF7, 0xA5B38140 */
148 w0  =  4.18938533204672725052e-01, /* 0x3FDACFE3, 0x90C97D69 */
149 w1  =  8.33333333333329678849e-02, /* 0x3FB55555, 0x5555553B */
150 w2  = -2.77777777728775536470e-03, /* 0xBF66C16C, 0x16B02E5C */
151 w3  =  7.93650558643019558500e-04, /* 0x3F4A019F, 0x98CF38B6 */
152 w4  = -5.95187557450339963135e-04, /* 0xBF4380CB, 0x8C0FE741 */
153 w5  =  8.36339918996282139126e-04, /* 0x3F4B67BA, 0x4CDAD5D1 */
154 w6  = -1.63092934096575273989e-03; /* 0xBF5AB89D, 0x0B9E43E4 */
155 
156 /*
157  * Compute sin(pi*x) without actually doing the pi*x multiplication.
158  * sin_pi(x) is only called for x < 0 and |x| < 2**(p-1) where p is
159  * the precision of x.
160  */
161 static double
162 sin_pi(double x)
163 {
164 	volatile double vz;
165 	double y,z;
166 	int n;
167 
168 	y = -x;
169 
170 	vz = y+0x1p52;			/* depend on 0 <= y < 0x1p52 */
171 	z = vz-0x1p52;			/* rint(y) for the above range */
172 	if (z == y)
173 	    return zero;
174 
175 	vz = y+0x1p50;
176 	GET_LOW_WORD(n,vz);		/* bits for rounded y (units 0.25) */
177 	z = vz-0x1p50;			/* y rounded to a multiple of 0.25 */
178 	if (z > y) {
179 	    z -= 0.25;			/* adjust to round down */
180 	    n--;
181 	}
182 	n &= 7;				/* octant of y mod 2 */
183 	y = y - z + n * 0.25;		/* y mod 2 */
184 
185 	switch (n) {
186 	    case 0:   y =  __kernel_sin(pi*y,zero,0); break;
187 	    case 1:
188 	    case 2:   y =  __kernel_cos(pi*(0.5-y),zero); break;
189 	    case 3:
190 	    case 4:   y =  __kernel_sin(pi*(one-y),zero,0); break;
191 	    case 5:
192 	    case 6:   y = -__kernel_cos(pi*(y-1.5),zero); break;
193 	    default:  y =  __kernel_sin(pi*(y-2.0),zero,0); break;
194 	    }
195 	return -y;
196 }
197 
198 
199 double
200 lgamma_r(double x, int *signgamp)
201 {
202 	double nadj,p,p1,p2,p3,q,r,t,w,y,z;
203 	int32_t hx;
204 	int i,ix,lx;
205 
206 	EXTRACT_WORDS(hx,lx,x);
207 
208     /* purge +-Inf and NaNs */
209 	*signgamp = 1;
210 	ix = hx&0x7fffffff;
211 	if(ix>=0x7ff00000) return x*x;
212 
213     /* purge +-0 and tiny arguments */
214 	*signgamp = 1-2*((uint32_t)hx>>31);
215 	if(ix<0x3c700000) {	/* |x|<2**-56, return -log(|x|) */
216 	    if((ix|lx)==0)
217 	        return one/vzero;
218 	    return -log(fabs(x));
219 	}
220 
221     /* purge negative integers and start evaluation for other x < 0 */
222 	if(hx<0) {
223 	    *signgamp = 1;
224 	    if(ix>=0x43300000) 	/* |x|>=2**52, must be -integer */
225 		return one/vzero;
226 	    t = sin_pi(x);
227 	    if(t==zero) return one/vzero; /* -integer */
228 	    nadj = log(pi/fabs(t*x));
229 	    if(t<zero) *signgamp = -1;
230 	    x = -x;
231 	}
232 
233     /* purge 1 and 2 */
234 	if((((ix-0x3ff00000)|lx)==0)||(((ix-0x40000000)|lx)==0)) r = 0;
235     /* for x < 2.0 */
236 	else if(ix<0x40000000) {
237 	    if(ix<=0x3feccccc) { 	/* lgamma(x) = lgamma(x+1)-log(x) */
238 		r = -log(x);
239 		if(ix>=0x3FE76944) {y = one-x; i= 0;}
240 		else if(ix>=0x3FCDA661) {y= x-(tc-one); i=1;}
241 	  	else {y = x; i=2;}
242 	    } else {
243 	  	r = zero;
244 	        if(ix>=0x3FFBB4C3) {y=2.0-x;i=0;} /* [1.7316,2] */
245 	        else if(ix>=0x3FF3B4C4) {y=x-tc;i=1;} /* [1.23,1.73] */
246 		else {y=x-one;i=2;}
247 	    }
248 	    switch(i) {
249 	      case 0:
250 		z = y*y;
251 		p1 = a0+z*(a2+z*(a4+z*(a6+z*(a8+z*a10))));
252 		p2 = z*(a1+z*(a3+z*(a5+z*(a7+z*(a9+z*a11)))));
253 		p  = y*p1+p2;
254 		r  += p-y/2; break;
255 	      case 1:
256 		z = y*y;
257 		w = z*y;
258 		p1 = t0+w*(t3+w*(t6+w*(t9 +w*t12)));	/* parallel comp */
259 		p2 = t1+w*(t4+w*(t7+w*(t10+w*t13)));
260 		p3 = t2+w*(t5+w*(t8+w*(t11+w*t14)));
261 		p  = z*p1-(tt-w*(p2+y*p3));
262 		r += tf + p; break;
263 	      case 2:
264 		p1 = y*(u0+y*(u1+y*(u2+y*(u3+y*(u4+y*u5)))));
265 		p2 = one+y*(v1+y*(v2+y*(v3+y*(v4+y*v5))));
266 		r += p1/p2-y/2;
267 	    }
268 	}
269     /* x < 8.0 */
270 	else if(ix<0x40200000) {
271 	    i = x;
272 	    y = x-i;
273 	    p = y*(s0+y*(s1+y*(s2+y*(s3+y*(s4+y*(s5+y*s6))))));
274 	    q = one+y*(r1+y*(r2+y*(r3+y*(r4+y*(r5+y*r6)))));
275 	    r = y/2+p/q;
276 	    z = one;	/* lgamma(1+s) = log(s) + lgamma(s) */
277 	    switch(i) {
278 	    case 7: z *= (y+6);		/* FALLTHRU */
279 	    case 6: z *= (y+5);		/* FALLTHRU */
280 	    case 5: z *= (y+4);		/* FALLTHRU */
281 	    case 4: z *= (y+3);		/* FALLTHRU */
282 	    case 3: z *= (y+2);		/* FALLTHRU */
283 		    r += log(z); break;
284 	    }
285     /* 8.0 <= x < 2**56 */
286 	} else if (ix < 0x43700000) {
287 	    t = log(x);
288 	    z = one/x;
289 	    y = z*z;
290 	    w = w0+z*(w1+y*(w2+y*(w3+y*(w4+y*(w5+y*w6)))));
291 	    r = (x-half)*(t-one)+w;
292 	} else
293     /* 2**56 <= x <= inf */
294 	    r =  x*(log(x)-one);
295 	if(hx<0) r = nadj - r;
296 	return r;
297 }
298 
299 #if (LDBL_MANT_DIG == 53)
300 __weak_reference(lgamma_r, lgammal_r);
301 #endif
302