xref: /openbsd/lib/libm/src/b_tgamma.c (revision 2f2c0062)
1 /*	$OpenBSD: b_tgamma.c,v 1.10 2016/09/12 19:47:02 guenther Exp $	*/
2 /*-
3  * Copyright (c) 1992, 1993
4  *	The Regents of the University of California.  All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 3. Neither the name of the University nor the names of its contributors
15  *    may be used to endorse or promote products derived from this software
16  *    without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 /*
32  * This code by P. McIlroy, Oct 1992;
33  *
34  * The financial support of UUNET Communications Services is greatfully
35  * acknowledged.
36  */
37 
38 #include <float.h>
39 #include <math.h>
40 
41 #include "math_private.h"
42 
43 /* METHOD:
44  * x < 0: Use reflection formula, G(x) = pi/(sin(pi*x)*x*G(x))
45  *	At negative integers, return NaN and raise invalid.
46  *
47  * x < 6.5:
48  *	Use argument reduction G(x+1) = xG(x) to reach the
49  *	range [1.066124,2.066124].  Use a rational
50  *	approximation centered at the minimum (x0+1) to
51  *	ensure monotonicity.
52  *
53  * x >= 6.5: Use the asymptotic approximation (Stirling's formula)
54  *	adjusted for equal-ripples:
55  *
56  *	log(G(x)) ~= (x-.5)*(log(x)-1) + .5(log(2*pi)-1) + 1/x*P(1/(x*x))
57  *
58  *	Keep extra precision in multiplying (x-.5)(log(x)-1), to
59  *	avoid premature round-off.
60  *
61  * Special values:
62  *	-Inf:			return NaN and raise invalid;
63  *	negative integer:	return NaN and raise invalid;
64  *	other x ~< -177.79:	return +-0 and raise underflow;
65  *	+-0:			return +-Inf and raise divide-by-zero;
66  *	finite x ~> 171.63:	return +Inf and raise overflow;
67  *	+Inf:			return +Inf;
68  *	NaN: 			return NaN.
69  *
70  * Accuracy: tgamma(x) is accurate to within
71  *	x > 0:  error provably < 0.9ulp.
72  *	Maximum observed in 1,000,000 trials was .87ulp.
73  *	x < 0:
74  *	Maximum observed error < 4ulp in 1,000,000 trials.
75  */
76 
77 static double neg_gam(double);
78 static double small_gam(double);
79 static double smaller_gam(double);
80 static struct Double large_gam(double);
81 static struct Double ratfun_gam(double, double);
82 
83 /*
84  * Rational approximation, A0 + x*x*P(x)/Q(x), on the interval
85  * [1.066.., 2.066..] accurate to 4.25e-19.
86  */
87 #define LEFT -.3955078125	/* left boundary for rat. approx */
88 #define x0 .461632144968362356785	/* xmin - 1 */
89 
90 #define a0_hi 0.88560319441088874992
91 #define a0_lo -.00000000000000004996427036469019695
92 #define P0	 6.21389571821820863029017800727e-01
93 #define P1	 2.65757198651533466104979197553e-01
94 #define P2	 5.53859446429917461063308081748e-03
95 #define P3	 1.38456698304096573887145282811e-03
96 #define P4	 2.40659950032711365819348969808e-03
97 #define Q0	 1.45019531250000000000000000000e+00
98 #define Q1	 1.06258521948016171343454061571e+00
99 #define Q2	-2.07474561943859936441469926649e-01
100 #define Q3	-1.46734131782005422506287573015e-01
101 #define Q4	 3.07878176156175520361557573779e-02
102 #define Q5	 5.12449347980666221336054633184e-03
103 #define Q6	-1.76012741431666995019222898833e-03
104 #define Q7	 9.35021023573788935372153030556e-05
105 #define Q8	 6.13275507472443958924745652239e-06
106 /*
107  * Constants for large x approximation (x in [6, Inf])
108  * (Accurate to 2.8*10^-19 absolute)
109  */
110 #define lns2pi_hi 0.418945312500000
111 #define lns2pi_lo -.000006779295327258219670263595
112 #define Pa0	 8.33333333333333148296162562474e-02
113 #define Pa1	-2.77777777774548123579378966497e-03
114 #define Pa2	 7.93650778754435631476282786423e-04
115 #define Pa3	-5.95235082566672847950717262222e-04
116 #define Pa4	 8.41428560346653702135821806252e-04
117 #define Pa5	-1.89773526463879200348872089421e-03
118 #define Pa6	 5.69394463439411649408050664078e-03
119 #define Pa7	-1.44705562421428915453880392761e-02
120 
121 static const double zero = 0., one = 1.0, tiny = 1e-300;
122 
123 double
tgamma(double x)124 tgamma(double x)
125 {
126 	struct Double u;
127 
128 	if (x >= 6) {
129 		if(x > 171.63)
130 			return(x/zero);
131 		u = large_gam(x);
132 		return(__exp__D(u.a, u.b));
133 	} else if (x >= 1.0 + LEFT + x0)
134 		return (small_gam(x));
135 	else if (x > 1.e-17)
136 		return (smaller_gam(x));
137 	else if (x > -1.e-17) {
138 		if (x != 0.0)
139 			u.a = one - tiny;	/* raise inexact */
140 		return (one/x);
141 	} else if (!isfinite(x)) {
142 		return (x - x);			/* x = NaN, -Inf */
143 	 } else
144 		return (neg_gam(x));
145 }
146 DEF_STD(tgamma);
147 LDBL_MAYBE_UNUSED_CLONE(tgamma);
148 
149 /*
150  * We simply call tgamma() rather than bloating the math library
151  * with a float-optimized version of it.  The reason is that tgammaf()
152  * is essentially useless, since the function is superexponential
153  * and floats have very limited range.  -- das@freebsd.org
154  */
155 
156 float
tgammaf(float x)157 tgammaf(float x)
158 {
159 	return tgamma(x);
160 }
161 
162 /*
163  * Accurate to max(ulp(1/128) absolute, 2^-66 relative) error.
164  */
165 
166 static struct Double
large_gam(double x)167 large_gam(double x)
168 {
169 	double z, p;
170 	struct Double t, u, v;
171 
172 	z = one/(x*x);
173 	p = Pa0+z*(Pa1+z*(Pa2+z*(Pa3+z*(Pa4+z*(Pa5+z*(Pa6+z*Pa7))))));
174 	p = p/x;
175 
176 	u = __log__D(x);
177 	u.a -= one;
178 	v.a = (x -= .5);
179 	TRUNC(v.a);
180 	v.b = x - v.a;
181 	t.a = v.a*u.a;			/* t = (x-.5)*(log(x)-1) */
182 	t.b = v.b*u.a + x*u.b;
183 	/* return t.a + t.b + lns2pi_hi + lns2pi_lo + p */
184 	t.b += lns2pi_lo; t.b += p;
185 	u.a = lns2pi_hi + t.b; u.a += t.a;
186 	u.b = t.a - u.a;
187 	u.b += lns2pi_hi; u.b += t.b;
188 	return (u);
189 }
190 
191 /*
192  * Good to < 1 ulp.  (provably .90 ulp; .87 ulp on 1,000,000 runs.)
193  * It also has correct monotonicity.
194  */
195 
196 static double
small_gam(double x)197 small_gam(double x)
198 {
199 	double y, ym1, t;
200 	struct Double yy, r;
201 	y = x - one;
202 	ym1 = y - one;
203 	if (y <= 1.0 + (LEFT + x0)) {
204 		yy = ratfun_gam(y - x0, 0);
205 		return (yy.a + yy.b);
206 	}
207 	r.a = y;
208 	TRUNC(r.a);
209 	yy.a = r.a - one;
210 	y = ym1;
211 	yy.b = r.b = y - yy.a;
212 	/* Argument reduction: G(x+1) = x*G(x) */
213 	for (ym1 = y-one; ym1 > LEFT + x0; y = ym1--, yy.a--) {
214 		t = r.a*yy.a;
215 		r.b = r.a*yy.b + y*r.b;
216 		r.a = t;
217 		TRUNC(r.a);
218 		r.b += (t - r.a);
219 	}
220 	/* Return r*tgamma(y). */
221 	yy = ratfun_gam(y - x0, 0);
222 	y = r.b*(yy.a + yy.b) + r.a*yy.b;
223 	y += yy.a*r.a;
224 	return (y);
225 }
226 
227 /*
228  * Good on (0, 1+x0+LEFT].  Accurate to 1ulp.
229  */
230 
231 static double
smaller_gam(double x)232 smaller_gam(double x)
233 {
234 	double t, d;
235 	struct Double r, xx;
236 	if (x < x0 + LEFT) {
237 		t = x;
238 		TRUNC(t);
239 		d = (t+x)*(x-t);
240 		t *= t;
241 		xx.a = (t + x);
242 		TRUNC(xx.a);
243 		xx.b = x - xx.a; xx.b += t; xx.b += d;
244 		t = (one-x0); t += x;
245 		d = (one-x0); d -= t; d += x;
246 		x = xx.a + xx.b;
247 	} else {
248 		xx.a =  x;
249 		TRUNC(xx.a);
250 		xx.b = x - xx.a;
251 		t = x - x0;
252 		d = (-x0 -t); d += x;
253 	}
254 	r = ratfun_gam(t, d);
255 	d = r.a/x;
256 	TRUNC(d);
257 	r.a -= d*xx.a; r.a -= d*xx.b; r.a += r.b;
258 	return (d + r.a/x);
259 }
260 
261 /*
262  * returns (z+c)^2 * P(z)/Q(z) + a0
263  */
264 
265 static struct Double
ratfun_gam(double z,double c)266 ratfun_gam(double z, double c)
267 {
268 	double p, q;
269 	struct Double r, t;
270 
271 	q = Q0 +z*(Q1+z*(Q2+z*(Q3+z*(Q4+z*(Q5+z*(Q6+z*(Q7+z*Q8)))))));
272 	p = P0 + z*(P1 + z*(P2 + z*(P3 + z*P4)));
273 
274 	/* return r.a + r.b = a0 + (z+c)^2*p/q, with r.a truncated to 26 bits. */
275 	p = p/q;
276 	t.a = z;
277 	TRUNC(t.a);			/* t ~= z + c */
278 	t.b = (z - t.a) + c;
279 	t.b *= (t.a + z);
280 	q = (t.a *= t.a);		/* t = (z+c)^2 */
281 	TRUNC(t.a);
282 	t.b += (q - t.a);
283 	r.a = p;
284 	TRUNC(r.a);			/* r = P/Q */
285 	r.b = p - r.a;
286 	t.b = t.b*p + t.a*r.b + a0_lo;
287 	t.a *= r.a;			/* t = (z+c)^2*(P/Q) */
288 	r.a = t.a + a0_hi;
289 	TRUNC(r.a);
290 	r.b = ((a0_hi-r.a) + t.a) + t.b;
291 	return (r);			/* r = a0 + t */
292 }
293 
294 static double
neg_gam(double x)295 neg_gam(double x)
296 {
297 	int sgn = 1;
298 	struct Double lg, lsine;
299 	double y, z;
300 
301 	y = ceil(x);
302 	if (y == x)		/* Negative integer. */
303 		return ((x - x) / zero);
304 	z = y - x;
305 	if (z > 0.5)
306 		z = one - z;
307 	y = 0.5 * y;
308 	if (y == ceil(y))
309 		sgn = -1;
310 	if (z < .25)
311 		z = sin(M_PI*z);
312 	else
313 		z = cos(M_PI*(0.5-z));
314 	/* Special case: G(1-x) = Inf; G(x) may be nonzero. */
315 	if (x < -170) {
316 		if (x < -190)
317 			return ((double)sgn*tiny*tiny);
318 		y = one - x;		/* exact: 128 < |x| < 255 */
319 		lg = large_gam(y);
320 		lsine = __log__D(M_PI/z);	/* = TRUNC(log(u)) + small */
321 		lg.a -= lsine.a;		/* exact (opposite signs) */
322 		lg.b -= lsine.b;
323 		y = -(lg.a + lg.b);
324 		z = (y + lg.a) + lg.b;
325 		y = __exp__D(y, z);
326 		if (sgn < 0) y = -y;
327 		return (y);
328 	}
329 	y = one-x;
330 	if (one-y == x)
331 		y = tgamma(y);
332 	else		/* 1-x is inexact */
333 		y = -x*tgamma(-x);
334 	if (sgn < 0) y = -y;
335 	return (M_PI / (y*z));
336 }
337