sync
[bitrig.git] / lib / libm / src / b_tgamma.c
blobfb5a4b369940ff30d0e2821c2974a0564a13a4a7
1 /* $OpenBSD: b_tgamma.c,v 1.7 2013/03/28 18:09:38 martynas Exp $ */
2 /*-
3 * Copyright (c) 1992, 1993
4 * The Regents of the University of California. All rights reserved.
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.
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.
32 * This code by P. McIlroy, Oct 1992;
34 * The financial support of UUNET Communications Services is greatfully
35 * acknowledged.
38 #include <float.h>
39 #include <math.h>
41 #include "math_private.h"
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.
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.
53 * x >= 6.5: Use the asymptotic approximation (Stirling's formula)
54 * adjusted for equal-ripples:
56 * log(G(x)) ~= (x-.5)*(log(x)-1) + .5(log(2*pi)-1) + 1/x*P(1/(x*x))
58 * Keep extra precision in multiplying (x-.5)(log(x)-1), to
59 * avoid premature round-off.
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.
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.
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);
84 * Rational approximation, A0 + x*x*P(x)/Q(x), on the interval
85 * [1.066.., 2.066..] accurate to 4.25e-19.
87 #define LEFT -.3955078125 /* left boundary for rat. approx */
88 #define x0 .461632144968362356785 /* xmin - 1 */
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
107 * Constants for large x approximation (x in [6, Inf])
108 * (Accurate to 2.8*10^-19 absolute)
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
121 static const double zero = 0., one = 1.0, tiny = 1e-300;
123 double
124 tgamma(double x)
126 struct Double u;
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 (!finite(x)) {
142 return (x - x); /* x = NaN, -Inf */
143 } else
144 return (neg_gam(x));
148 * We simply call tgamma() rather than bloating the math library
149 * with a float-optimized version of it. The reason is that tgammaf()
150 * is essentially useless, since the function is superexponential
151 * and floats have very limited range. -- das@freebsd.org
154 float
155 tgammaf(float x)
157 return tgamma(x);
161 * Accurate to max(ulp(1/128) absolute, 2^-66 relative) error.
164 static struct Double
165 large_gam(double x)
167 double z, p;
168 struct Double t, u, v;
170 z = one/(x*x);
171 p = Pa0+z*(Pa1+z*(Pa2+z*(Pa3+z*(Pa4+z*(Pa5+z*(Pa6+z*Pa7))))));
172 p = p/x;
174 u = __log__D(x);
175 u.a -= one;
176 v.a = (x -= .5);
177 TRUNC(v.a);
178 v.b = x - v.a;
179 t.a = v.a*u.a; /* t = (x-.5)*(log(x)-1) */
180 t.b = v.b*u.a + x*u.b;
181 /* return t.a + t.b + lns2pi_hi + lns2pi_lo + p */
182 t.b += lns2pi_lo; t.b += p;
183 u.a = lns2pi_hi + t.b; u.a += t.a;
184 u.b = t.a - u.a;
185 u.b += lns2pi_hi; u.b += t.b;
186 return (u);
190 * Good to < 1 ulp. (provably .90 ulp; .87 ulp on 1,000,000 runs.)
191 * It also has correct monotonicity.
194 static double
195 small_gam(double x)
197 double y, ym1, t;
198 struct Double yy, r;
199 y = x - one;
200 ym1 = y - one;
201 if (y <= 1.0 + (LEFT + x0)) {
202 yy = ratfun_gam(y - x0, 0);
203 return (yy.a + yy.b);
205 r.a = y;
206 TRUNC(r.a);
207 yy.a = r.a - one;
208 y = ym1;
209 yy.b = r.b = y - yy.a;
210 /* Argument reduction: G(x+1) = x*G(x) */
211 for (ym1 = y-one; ym1 > LEFT + x0; y = ym1--, yy.a--) {
212 t = r.a*yy.a;
213 r.b = r.a*yy.b + y*r.b;
214 r.a = t;
215 TRUNC(r.a);
216 r.b += (t - r.a);
218 /* Return r*tgamma(y). */
219 yy = ratfun_gam(y - x0, 0);
220 y = r.b*(yy.a + yy.b) + r.a*yy.b;
221 y += yy.a*r.a;
222 return (y);
226 * Good on (0, 1+x0+LEFT]. Accurate to 1ulp.
229 static double
230 smaller_gam(double x)
232 double t, d;
233 struct Double r, xx;
234 if (x < x0 + LEFT) {
235 t = x;
236 TRUNC(t);
237 d = (t+x)*(x-t);
238 t *= t;
239 xx.a = (t + x);
240 TRUNC(xx.a);
241 xx.b = x - xx.a; xx.b += t; xx.b += d;
242 t = (one-x0); t += x;
243 d = (one-x0); d -= t; d += x;
244 x = xx.a + xx.b;
245 } else {
246 xx.a = x;
247 TRUNC(xx.a);
248 xx.b = x - xx.a;
249 t = x - x0;
250 d = (-x0 -t); d += x;
252 r = ratfun_gam(t, d);
253 d = r.a/x;
254 TRUNC(d);
255 r.a -= d*xx.a; r.a -= d*xx.b; r.a += r.b;
256 return (d + r.a/x);
260 * returns (z+c)^2 * P(z)/Q(z) + a0
263 static struct Double
264 ratfun_gam(double z, double c)
266 double p, q;
267 struct Double r, t;
269 q = Q0 +z*(Q1+z*(Q2+z*(Q3+z*(Q4+z*(Q5+z*(Q6+z*(Q7+z*Q8)))))));
270 p = P0 + z*(P1 + z*(P2 + z*(P3 + z*P4)));
272 /* return r.a + r.b = a0 + (z+c)^2*p/q, with r.a truncated to 26 bits. */
273 p = p/q;
274 t.a = z;
275 TRUNC(t.a); /* t ~= z + c */
276 t.b = (z - t.a) + c;
277 t.b *= (t.a + z);
278 q = (t.a *= t.a); /* t = (z+c)^2 */
279 TRUNC(t.a);
280 t.b += (q - t.a);
281 r.a = p;
282 TRUNC(r.a); /* r = P/Q */
283 r.b = p - r.a;
284 t.b = t.b*p + t.a*r.b + a0_lo;
285 t.a *= r.a; /* t = (z+c)^2*(P/Q) */
286 r.a = t.a + a0_hi;
287 TRUNC(r.a);
288 r.b = ((a0_hi-r.a) + t.a) + t.b;
289 return (r); /* r = a0 + t */
292 static double
293 neg_gam(double x)
295 int sgn = 1;
296 struct Double lg, lsine;
297 double y, z;
299 y = ceil(x);
300 if (y == x) /* Negative integer. */
301 return ((x - x) / zero);
302 z = y - x;
303 if (z > 0.5)
304 z = one - z;
305 y = 0.5 * y;
306 if (y == ceil(y))
307 sgn = -1;
308 if (z < .25)
309 z = sin(M_PI*z);
310 else
311 z = cos(M_PI*(0.5-z));
312 /* Special case: G(1-x) = Inf; G(x) may be nonzero. */
313 if (x < -170) {
314 if (x < -190)
315 return ((double)sgn*tiny*tiny);
316 y = one - x; /* exact: 128 < |x| < 255 */
317 lg = large_gam(y);
318 lsine = __log__D(M_PI/z); /* = TRUNC(log(u)) + small */
319 lg.a -= lsine.a; /* exact (opposite signs) */
320 lg.b -= lsine.b;
321 y = -(lg.a + lg.b);
322 z = (y + lg.a) + lg.b;
323 y = __exp__D(y, z);
324 if (sgn < 0) y = -y;
325 return (y);
327 y = one-x;
328 if (one-y == x)
329 y = tgamma(y);
330 else /* 1-x is inexact */
331 y = -x*tgamma(-x);
332 if (sgn < 0) y = -y;
333 return (M_PI / (y*z));
336 #if LDBL_MANT_DIG == 53
337 __strong_alias(tgammal, tgamma);
338 #endif /* LDBL_MANT_DIG == 53 */