xref: /titanic_51/usr/src/lib/libast/common/uwin/gamma.c (revision da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968)
1*da2e3ebdSchin #include "FEATURE/uwin"
2*da2e3ebdSchin 
3*da2e3ebdSchin #if !_UWIN || _lib_gamma
4*da2e3ebdSchin 
_STUB_gamma()5*da2e3ebdSchin void _STUB_gamma(){}
6*da2e3ebdSchin 
7*da2e3ebdSchin #else
8*da2e3ebdSchin 
9*da2e3ebdSchin /*-
10*da2e3ebdSchin  * Copyright (c) 1992, 1993
11*da2e3ebdSchin  *	The Regents of the University of California.  All rights reserved.
12*da2e3ebdSchin  *
13*da2e3ebdSchin  * Redistribution and use in source and binary forms, with or without
14*da2e3ebdSchin  * modification, are permitted provided that the following conditions
15*da2e3ebdSchin  * are met:
16*da2e3ebdSchin  * 1. Redistributions of source code must retain the above copyright
17*da2e3ebdSchin  *    notice, this list of conditions and the following disclaimer.
18*da2e3ebdSchin  * 2. Redistributions in binary form must reproduce the above copyright
19*da2e3ebdSchin  *    notice, this list of conditions and the following disclaimer in the
20*da2e3ebdSchin  *    documentation and/or other materials provided with the distribution.
21*da2e3ebdSchin  * 3. Neither the name of the University nor the names of its contributors
22*da2e3ebdSchin  *    may be used to endorse or promote products derived from this software
23*da2e3ebdSchin  *    without specific prior written permission.
24*da2e3ebdSchin  *
25*da2e3ebdSchin  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
26*da2e3ebdSchin  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27*da2e3ebdSchin  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28*da2e3ebdSchin  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
29*da2e3ebdSchin  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30*da2e3ebdSchin  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31*da2e3ebdSchin  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32*da2e3ebdSchin  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33*da2e3ebdSchin  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34*da2e3ebdSchin  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35*da2e3ebdSchin  * SUCH DAMAGE.
36*da2e3ebdSchin  */
37*da2e3ebdSchin 
38*da2e3ebdSchin #ifndef lint
39*da2e3ebdSchin static char sccsid[] = "@(#)gamma.c	8.1 (Berkeley) 6/4/93";
40*da2e3ebdSchin #endif /* not lint */
41*da2e3ebdSchin 
42*da2e3ebdSchin /*
43*da2e3ebdSchin  * This code by P. McIlroy, Oct 1992;
44*da2e3ebdSchin  *
45*da2e3ebdSchin  * The financial support of UUNET Communications Services is greatfully
46*da2e3ebdSchin  * acknowledged.
47*da2e3ebdSchin  */
48*da2e3ebdSchin 
49*da2e3ebdSchin #define gamma	______gamma
50*da2e3ebdSchin 
51*da2e3ebdSchin #include <math.h>
52*da2e3ebdSchin #include <errno.h>
53*da2e3ebdSchin #include "mathimpl.h"
54*da2e3ebdSchin 
55*da2e3ebdSchin #undef	gamma
56*da2e3ebdSchin 
57*da2e3ebdSchin /* METHOD:
58*da2e3ebdSchin  * x < 0: Use reflection formula, G(x) = pi/(sin(pi*x)*x*G(x))
59*da2e3ebdSchin  * 	At negative integers, return +Inf, and set errno.
60*da2e3ebdSchin  *
61*da2e3ebdSchin  * x < 6.5:
62*da2e3ebdSchin  *	Use argument reduction G(x+1) = xG(x) to reach the
63*da2e3ebdSchin  *	range [1.066124,2.066124].  Use a rational
64*da2e3ebdSchin  *	approximation centered at the minimum (x0+1) to
65*da2e3ebdSchin  *	ensure monotonicity.
66*da2e3ebdSchin  *
67*da2e3ebdSchin  * x >= 6.5: Use the asymptotic approximation (Stirling's formula)
68*da2e3ebdSchin  *	adjusted for equal-ripples:
69*da2e3ebdSchin  *
70*da2e3ebdSchin  *	log(G(x)) ~= (x-.5)*(log(x)-1) + .5(log(2*pi)-1) + 1/x*P(1/(x*x))
71*da2e3ebdSchin  *
72*da2e3ebdSchin  *	Keep extra precision in multiplying (x-.5)(log(x)-1), to
73*da2e3ebdSchin  *	avoid premature round-off.
74*da2e3ebdSchin  *
75*da2e3ebdSchin  * Special values:
76*da2e3ebdSchin  *	non-positive integer:	Set overflow trap; return +Inf;
77*da2e3ebdSchin  *	x > 171.63:		Set overflow trap; return +Inf;
78*da2e3ebdSchin  *	NaN: 			Set invalid trap;  return NaN
79*da2e3ebdSchin  *
80*da2e3ebdSchin  * Accuracy: Gamma(x) is accurate to within
81*da2e3ebdSchin  *	x > 0:  error provably < 0.9ulp.
82*da2e3ebdSchin  *	Maximum observed in 1,000,000 trials was .87ulp.
83*da2e3ebdSchin  *	x < 0:
84*da2e3ebdSchin  *	Maximum observed error < 4ulp in 1,000,000 trials.
85*da2e3ebdSchin  */
86*da2e3ebdSchin 
87*da2e3ebdSchin static double neg_gam __P((double));
88*da2e3ebdSchin static double small_gam __P((double));
89*da2e3ebdSchin static double smaller_gam __P((double));
90*da2e3ebdSchin static struct Double large_gam __P((double));
91*da2e3ebdSchin static struct Double ratfun_gam __P((double, double));
92*da2e3ebdSchin 
93*da2e3ebdSchin /*
94*da2e3ebdSchin  * Rational approximation, A0 + x*x*P(x)/Q(x), on the interval
95*da2e3ebdSchin  * [1.066.., 2.066..] accurate to 4.25e-19.
96*da2e3ebdSchin  */
97*da2e3ebdSchin #define LEFT -.3955078125	/* left boundary for rat. approx */
98*da2e3ebdSchin #define x0 .461632144968362356785	/* xmin - 1 */
99*da2e3ebdSchin 
100*da2e3ebdSchin #define a0_hi 0.88560319441088874992
101*da2e3ebdSchin #define a0_lo -.00000000000000004996427036469019695
102*da2e3ebdSchin #define P0	 6.21389571821820863029017800727e-01
103*da2e3ebdSchin #define P1	 2.65757198651533466104979197553e-01
104*da2e3ebdSchin #define P2	 5.53859446429917461063308081748e-03
105*da2e3ebdSchin #define P3	 1.38456698304096573887145282811e-03
106*da2e3ebdSchin #define P4	 2.40659950032711365819348969808e-03
107*da2e3ebdSchin #define Q0	 1.45019531250000000000000000000e+00
108*da2e3ebdSchin #define Q1	 1.06258521948016171343454061571e+00
109*da2e3ebdSchin #define Q2	-2.07474561943859936441469926649e-01
110*da2e3ebdSchin #define Q3	-1.46734131782005422506287573015e-01
111*da2e3ebdSchin #define Q4	 3.07878176156175520361557573779e-02
112*da2e3ebdSchin #define Q5	 5.12449347980666221336054633184e-03
113*da2e3ebdSchin #define Q6	-1.76012741431666995019222898833e-03
114*da2e3ebdSchin #define Q7	 9.35021023573788935372153030556e-05
115*da2e3ebdSchin #define Q8	 6.13275507472443958924745652239e-06
116*da2e3ebdSchin /*
117*da2e3ebdSchin  * Constants for large x approximation (x in [6, Inf])
118*da2e3ebdSchin  * (Accurate to 2.8*10^-19 absolute)
119*da2e3ebdSchin  */
120*da2e3ebdSchin #define lns2pi_hi 0.418945312500000
121*da2e3ebdSchin #define lns2pi_lo -.000006779295327258219670263595
122*da2e3ebdSchin #define Pa0	 8.33333333333333148296162562474e-02
123*da2e3ebdSchin #define Pa1	-2.77777777774548123579378966497e-03
124*da2e3ebdSchin #define Pa2	 7.93650778754435631476282786423e-04
125*da2e3ebdSchin #define Pa3	-5.95235082566672847950717262222e-04
126*da2e3ebdSchin #define Pa4	 8.41428560346653702135821806252e-04
127*da2e3ebdSchin #define Pa5	-1.89773526463879200348872089421e-03
128*da2e3ebdSchin #define Pa6	 5.69394463439411649408050664078e-03
129*da2e3ebdSchin #define Pa7	-1.44705562421428915453880392761e-02
130*da2e3ebdSchin 
131*da2e3ebdSchin static const double zero = 0., one = 1.0, tiny = 1e-300;
132*da2e3ebdSchin static int endian;
133*da2e3ebdSchin /*
134*da2e3ebdSchin  * TRUNC sets trailing bits in a floating-point number to zero.
135*da2e3ebdSchin  * is a temporary variable.
136*da2e3ebdSchin  */
137*da2e3ebdSchin #if defined(vax) || defined(tahoe)
138*da2e3ebdSchin #define _IEEE		0
139*da2e3ebdSchin #define TRUNC(x)	x = (double) (float) (x)
140*da2e3ebdSchin #else
141*da2e3ebdSchin #define _IEEE		1
142*da2e3ebdSchin #define TRUNC(x)	*(((int *) &x) + endian) &= 0xf8000000
143*da2e3ebdSchin #define infnan(x)	0.0
144*da2e3ebdSchin #endif
145*da2e3ebdSchin 
146*da2e3ebdSchin extern double gamma(x)
147*da2e3ebdSchin 	double x;
148*da2e3ebdSchin {
149*da2e3ebdSchin 	struct Double u;
150*da2e3ebdSchin 	endian = (*(int *) &one) ? 1 : 0;
151*da2e3ebdSchin 
152*da2e3ebdSchin 	if (x >= 6) {
153*da2e3ebdSchin 		if(x > 171.63)
154*da2e3ebdSchin 			return(one/zero);
155*da2e3ebdSchin 		u = large_gam(x);
156*da2e3ebdSchin 		return(__exp__D(u.a, u.b));
157*da2e3ebdSchin 	} else if (x >= 1.0 + LEFT + x0)
158*da2e3ebdSchin 		return (small_gam(x));
159*da2e3ebdSchin 	else if (x > 1.e-17)
160*da2e3ebdSchin 		return (smaller_gam(x));
161*da2e3ebdSchin 	else if (x > -1.e-17) {
162*da2e3ebdSchin 		if (x == 0.0)
163*da2e3ebdSchin 			if (!_IEEE) return (infnan(ERANGE));
164*da2e3ebdSchin 			else return (one/x);
165*da2e3ebdSchin 		one+1e-20;		/* Raise inexact flag. */
166*da2e3ebdSchin 		return (one/x);
167*da2e3ebdSchin 	} else if (!finite(x)) {
168*da2e3ebdSchin 		if (_IEEE)		/* x = NaN, -Inf */
169*da2e3ebdSchin 			return (x*x);
170*da2e3ebdSchin 		else
171*da2e3ebdSchin 			return (infnan(EDOM));
172*da2e3ebdSchin 	 } else
173*da2e3ebdSchin 		return (neg_gam(x));
174*da2e3ebdSchin }
175*da2e3ebdSchin /*
176*da2e3ebdSchin  * Accurate to max(ulp(1/128) absolute, 2^-66 relative) error.
177*da2e3ebdSchin  */
178*da2e3ebdSchin static struct Double
large_gam(x)179*da2e3ebdSchin large_gam(x)
180*da2e3ebdSchin 	double x;
181*da2e3ebdSchin {
182*da2e3ebdSchin 	double z, p;
183*da2e3ebdSchin 	struct Double t, u, v;
184*da2e3ebdSchin 
185*da2e3ebdSchin 	z = one/(x*x);
186*da2e3ebdSchin 	p = Pa0+z*(Pa1+z*(Pa2+z*(Pa3+z*(Pa4+z*(Pa5+z*(Pa6+z*Pa7))))));
187*da2e3ebdSchin 	p = p/x;
188*da2e3ebdSchin 
189*da2e3ebdSchin 	u = __log__D(x);
190*da2e3ebdSchin 	u.a -= one;
191*da2e3ebdSchin 	v.a = (x -= .5);
192*da2e3ebdSchin 	TRUNC(v.a);
193*da2e3ebdSchin 	v.b = x - v.a;
194*da2e3ebdSchin 	t.a = v.a*u.a;			/* t = (x-.5)*(log(x)-1) */
195*da2e3ebdSchin 	t.b = v.b*u.a + x*u.b;
196*da2e3ebdSchin 	/* return t.a + t.b + lns2pi_hi + lns2pi_lo + p */
197*da2e3ebdSchin 	t.b += lns2pi_lo; t.b += p;
198*da2e3ebdSchin 	u.a = lns2pi_hi + t.b; u.a += t.a;
199*da2e3ebdSchin 	u.b = t.a - u.a;
200*da2e3ebdSchin 	u.b += lns2pi_hi; u.b += t.b;
201*da2e3ebdSchin 	return (u);
202*da2e3ebdSchin }
203*da2e3ebdSchin /*
204*da2e3ebdSchin  * Good to < 1 ulp.  (provably .90 ulp; .87 ulp on 1,000,000 runs.)
205*da2e3ebdSchin  * It also has correct monotonicity.
206*da2e3ebdSchin  */
207*da2e3ebdSchin static double
small_gam(x)208*da2e3ebdSchin small_gam(x)
209*da2e3ebdSchin 	double x;
210*da2e3ebdSchin {
211*da2e3ebdSchin 	double y, ym1, t;
212*da2e3ebdSchin 	struct Double yy, r;
213*da2e3ebdSchin 	y = x - one;
214*da2e3ebdSchin 	ym1 = y - one;
215*da2e3ebdSchin 	if (y <= 1.0 + (LEFT + x0)) {
216*da2e3ebdSchin 		yy = ratfun_gam(y - x0, 0);
217*da2e3ebdSchin 		return (yy.a + yy.b);
218*da2e3ebdSchin 	}
219*da2e3ebdSchin 	r.a = y;
220*da2e3ebdSchin 	TRUNC(r.a);
221*da2e3ebdSchin 	yy.a = r.a - one;
222*da2e3ebdSchin 	y = ym1;
223*da2e3ebdSchin 	yy.b = r.b = y - yy.a;
224*da2e3ebdSchin 	/* Argument reduction: G(x+1) = x*G(x) */
225*da2e3ebdSchin 	for (ym1 = y-one; ym1 > LEFT + x0; y = ym1--, yy.a--) {
226*da2e3ebdSchin 		t = r.a*yy.a;
227*da2e3ebdSchin 		r.b = r.a*yy.b + y*r.b;
228*da2e3ebdSchin 		r.a = t;
229*da2e3ebdSchin 		TRUNC(r.a);
230*da2e3ebdSchin 		r.b += (t - r.a);
231*da2e3ebdSchin 	}
232*da2e3ebdSchin 	/* Return r*gamma(y). */
233*da2e3ebdSchin 	yy = ratfun_gam(y - x0, 0);
234*da2e3ebdSchin 	y = r.b*(yy.a + yy.b) + r.a*yy.b;
235*da2e3ebdSchin 	y += yy.a*r.a;
236*da2e3ebdSchin 	return (y);
237*da2e3ebdSchin }
238*da2e3ebdSchin /*
239*da2e3ebdSchin  * Good on (0, 1+x0+LEFT].  Accurate to 1ulp.
240*da2e3ebdSchin  */
241*da2e3ebdSchin static double
smaller_gam(x)242*da2e3ebdSchin smaller_gam(x)
243*da2e3ebdSchin 	double x;
244*da2e3ebdSchin {
245*da2e3ebdSchin 	double t, d;
246*da2e3ebdSchin 	struct Double r, xx;
247*da2e3ebdSchin 	if (x < x0 + LEFT) {
248*da2e3ebdSchin 		t = x, TRUNC(t);
249*da2e3ebdSchin 		d = (t+x)*(x-t);
250*da2e3ebdSchin 		t *= t;
251*da2e3ebdSchin 		xx.a = (t + x), TRUNC(xx.a);
252*da2e3ebdSchin 		xx.b = x - xx.a; xx.b += t; xx.b += d;
253*da2e3ebdSchin 		t = (one-x0); t += x;
254*da2e3ebdSchin 		d = (one-x0); d -= t; d += x;
255*da2e3ebdSchin 		x = xx.a + xx.b;
256*da2e3ebdSchin 	} else {
257*da2e3ebdSchin 		xx.a =  x, TRUNC(xx.a);
258*da2e3ebdSchin 		xx.b = x - xx.a;
259*da2e3ebdSchin 		t = x - x0;
260*da2e3ebdSchin 		d = (-x0 -t); d += x;
261*da2e3ebdSchin 	}
262*da2e3ebdSchin 	r = ratfun_gam(t, d);
263*da2e3ebdSchin 	d = r.a/x, TRUNC(d);
264*da2e3ebdSchin 	r.a -= d*xx.a; r.a -= d*xx.b; r.a += r.b;
265*da2e3ebdSchin 	return (d + r.a/x);
266*da2e3ebdSchin }
267*da2e3ebdSchin /*
268*da2e3ebdSchin  * returns (z+c)^2 * P(z)/Q(z) + a0
269*da2e3ebdSchin  */
270*da2e3ebdSchin static struct Double
ratfun_gam(z,c)271*da2e3ebdSchin ratfun_gam(z, c)
272*da2e3ebdSchin 	double z, c;
273*da2e3ebdSchin {
274*da2e3ebdSchin 	double p, q;
275*da2e3ebdSchin 	struct Double r, t;
276*da2e3ebdSchin 
277*da2e3ebdSchin 	q = Q0 +z*(Q1+z*(Q2+z*(Q3+z*(Q4+z*(Q5+z*(Q6+z*(Q7+z*Q8)))))));
278*da2e3ebdSchin 	p = P0 + z*(P1 + z*(P2 + z*(P3 + z*P4)));
279*da2e3ebdSchin 
280*da2e3ebdSchin 	/* return r.a + r.b = a0 + (z+c)^2*p/q, with r.a truncated to 26 bits. */
281*da2e3ebdSchin 	p = p/q;
282*da2e3ebdSchin 	t.a = z, TRUNC(t.a);		/* t ~= z + c */
283*da2e3ebdSchin 	t.b = (z - t.a) + c;
284*da2e3ebdSchin 	t.b *= (t.a + z);
285*da2e3ebdSchin 	q = (t.a *= t.a);		/* t = (z+c)^2 */
286*da2e3ebdSchin 	TRUNC(t.a);
287*da2e3ebdSchin 	t.b += (q - t.a);
288*da2e3ebdSchin 	r.a = p, TRUNC(r.a);		/* r = P/Q */
289*da2e3ebdSchin 	r.b = p - r.a;
290*da2e3ebdSchin 	t.b = t.b*p + t.a*r.b + a0_lo;
291*da2e3ebdSchin 	t.a *= r.a;			/* t = (z+c)^2*(P/Q) */
292*da2e3ebdSchin 	r.a = t.a + a0_hi, TRUNC(r.a);
293*da2e3ebdSchin 	r.b = ((a0_hi-r.a) + t.a) + t.b;
294*da2e3ebdSchin 	return (r);			/* r = a0 + t */
295*da2e3ebdSchin }
296*da2e3ebdSchin 
297*da2e3ebdSchin static double
neg_gam(x)298*da2e3ebdSchin neg_gam(x)
299*da2e3ebdSchin 	double x;
300*da2e3ebdSchin {
301*da2e3ebdSchin 	int sgn = 1;
302*da2e3ebdSchin 	struct Double lg, lsine;
303*da2e3ebdSchin 	double y, z;
304*da2e3ebdSchin 
305*da2e3ebdSchin 	y = floor(x + .5);
306*da2e3ebdSchin 	if (y == x)		/* Negative integer. */
307*da2e3ebdSchin 		if(!_IEEE)
308*da2e3ebdSchin 			return (infnan(ERANGE));
309*da2e3ebdSchin 		else
310*da2e3ebdSchin 			return (one/zero);
311*da2e3ebdSchin 	z = fabs(x - y);
312*da2e3ebdSchin 	y = .5*ceil(x);
313*da2e3ebdSchin 	if (y == ceil(y))
314*da2e3ebdSchin 		sgn = -1;
315*da2e3ebdSchin 	if (z < .25)
316*da2e3ebdSchin 		z = sin(M_PI*z);
317*da2e3ebdSchin 	else
318*da2e3ebdSchin 		z = cos(M_PI*(0.5-z));
319*da2e3ebdSchin 	/* Special case: G(1-x) = Inf; G(x) may be nonzero. */
320*da2e3ebdSchin 	if (x < -170) {
321*da2e3ebdSchin 		if (x < -190)
322*da2e3ebdSchin 			return ((double)sgn*tiny*tiny);
323*da2e3ebdSchin 		y = one - x;		/* exact: 128 < |x| < 255 */
324*da2e3ebdSchin 		lg = large_gam(y);
325*da2e3ebdSchin 		lsine = __log__D(M_PI/z);	/* = TRUNC(log(u)) + small */
326*da2e3ebdSchin 		lg.a -= lsine.a;		/* exact (opposite signs) */
327*da2e3ebdSchin 		lg.b -= lsine.b;
328*da2e3ebdSchin 		y = -(lg.a + lg.b);
329*da2e3ebdSchin 		z = (y + lg.a) + lg.b;
330*da2e3ebdSchin 		y = __exp__D(y, z);
331*da2e3ebdSchin 		if (sgn < 0) y = -y;
332*da2e3ebdSchin 		return (y);
333*da2e3ebdSchin 	}
334*da2e3ebdSchin 	y = one-x;
335*da2e3ebdSchin 	if (one-y == x)
336*da2e3ebdSchin 		y = gamma(y);
337*da2e3ebdSchin 	else		/* 1-x is inexact */
338*da2e3ebdSchin 		y = -x*gamma(-x);
339*da2e3ebdSchin 	if (sgn < 0) y = -y;
340*da2e3ebdSchin 	return (M_PI / (y*z));
341*da2e3ebdSchin }
342*da2e3ebdSchin 
343*da2e3ebdSchin #endif
344