xref: /illumos-gate/usr/src/lib/libm/common/C/erf.c (revision f3041bfa)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #pragma weak __erf = erf
31 #pragma weak __erfc = erfc
32 
33 /* INDENT OFF */
34 /*
35  * double erf(double x)
36  * double erfc(double x)
37  *			     x
38  *		      2      |\
39  *     erf(x)  =  ---------  | exp(-t*t)dt
40  *		   sqrt(pi) \|
41  *			     0
42  *
43  *     erfc(x) =  1-erf(x)
44  *  Note that
45  *		erf(-x) = -erf(x)
46  *		erfc(-x) = 2 - erfc(x)
47  *
48  * Method:
49  *	1. For |x| in [0, 0.84375]
50  *	    erf(x)  = x + x*R(x^2)
51  *          erfc(x) = 1 - erf(x)           if x in [-.84375,0.25]
52  *                  = 0.5 + ((0.5-x)-x*R)  if x in [0.25,0.84375]
53  *	   where R = P/Q where P is an odd poly of degree 8 and
54  *	   Q is an odd poly of degree 10.
55  *						 -57.90
56  *			| R - (erf(x)-x)/x | <= 2
57  *
58  *
59  *	   Remark. The formula is derived by noting
60  *          erf(x) = (2/sqrt(pi))*(x - x^3/3 + x^5/10 - x^7/42 + ....)
61  *	   and that
62  *          2/sqrt(pi) = 1.128379167095512573896158903121545171688
63  *	   is close to one. The interval is chosen because the fix
64  *	   point of erf(x) is near 0.6174 (i.e., erf(x)=x when x is
65  *	   near 0.6174), and by some experiment, 0.84375 is chosen to
66  *	   guarantee the error is less than one ulp for erf.
67  *
68  *      2. For |x| in [0.84375,1.25], let s = |x| - 1, and
69  *         c = 0.84506291151 rounded to single (24 bits)
70  *         	erf(x)  = sign(x) * (c  + P1(s)/Q1(s))
71  *         	erfc(x) = (1-c)  - P1(s)/Q1(s) if x > 0
72  *			  1+(c+P1(s)/Q1(s))    if x < 0
73  *         	|P1/Q1 - (erf(|x|)-c)| <= 2**-59.06
74  *	   Remark: here we use the taylor series expansion at x=1.
75  *		erf(1+s) = erf(1) + s*Poly(s)
76  *			 = 0.845.. + P1(s)/Q1(s)
77  *	   That is, we use rational approximation to approximate
78  *			erf(1+s) - (c = (single)0.84506291151)
79  *	   Note that |P1/Q1|< 0.078 for x in [0.84375,1.25]
80  *	   where
81  *		P1(s) = degree 6 poly in s
82  *		Q1(s) = degree 6 poly in s
83  *
84  *      3. For x in [1.25,1/0.35(~2.857143)],
85  *         	erfc(x) = (1/x)*exp(-x*x-0.5625+R1/S1)
86  *         	erf(x)  = 1 - erfc(x)
87  *	   where
88  *		R1(z) = degree 7 poly in z, (z=1/x^2)
89  *		S1(z) = degree 8 poly in z
90  *
91  *      4. For x in [1/0.35,28]
92  *         	erfc(x) = (1/x)*exp(-x*x-0.5625+R2/S2) if x > 0
93  *			= 2.0 - (1/x)*exp(-x*x-0.5625+R2/S2) if -6<x<0
94  *			= 2.0 - tiny		(if x <= -6)
95  *         	erf(x)  = sign(x)*(1.0 - erfc(x)) if x < 6, else
96  *         	erf(x)  = sign(x)*(1.0 - tiny)
97  *	   where
98  *		R2(z) = degree 6 poly in z, (z=1/x^2)
99  *		S2(z) = degree 7 poly in z
100  *
101  *      Note1:
102  *	   To compute exp(-x*x-0.5625+R/S), let s be a single
103  *	   precision number and s := x; then
104  *		-x*x = -s*s + (s-x)*(s+x)
105  *	        exp(-x*x-0.5626+R/S) =
106  *			exp(-s*s-0.5625)*exp((s-x)*(s+x)+R/S);
107  *      Note2:
108  *	   Here 4 and 5 make use of the asymptotic series
109  *			  exp(-x*x)
110  *		erfc(x) ~ ---------- * ( 1 + Poly(1/x^2) )
111  *			  x*sqrt(pi)
112  *	   We use rational approximation to approximate
113  *      	g(s)=f(1/x^2) = log(erfc(x)*x) - x*x + 0.5625
114  *	   Here is the error bound for R1/S1 and R2/S2
115  *      	|R1/S1 - f(x)|  < 2**(-62.57)
116  *      	|R2/S2 - f(x)|  < 2**(-61.52)
117  *
118  *      5. For inf > x >= 28
119  *         	erf(x)  = sign(x) *(1 - tiny)  (raise inexact)
120  *         	erfc(x) = tiny*tiny (raise underflow) if x > 0
121  *			= 2 - tiny if x<0
122  *
123  *      7. Special case:
124  *         	erf(0)  = 0, erf(inf)  = 1, erf(-inf) = -1,
125  *         	erfc(0) = 1, erfc(inf) = 0, erfc(-inf) = 2,
126  *   	erfc/erf(NaN) is NaN
127  */
128 /* INDENT ON */
129 
130 #include "libm_macros.h"
131 #include <math.h>
132 
133 static const double xxx[] = {
134 /* tiny */	1e-300,
135 /* half */	5.00000000000000000000e-01,	/* 3FE00000, 00000000 */
136 /* one */	1.00000000000000000000e+00,	/* 3FF00000, 00000000 */
137 /* two */	2.00000000000000000000e+00,	/* 40000000, 00000000 */
138 /* erx */	8.45062911510467529297e-01,	/* 3FEB0AC1, 60000000 */
139 /*
140  * Coefficients for approximation to  erf on [0,0.84375]
141  */
142 /* efx */	 1.28379167095512586316e-01,	/* 3FC06EBA, 8214DB69 */
143 /* efx8 */	 1.02703333676410069053e+00,	/* 3FF06EBA, 8214DB69 */
144 /* pp0 */	 1.28379167095512558561e-01,	/* 3FC06EBA, 8214DB68 */
145 /* pp1 */	-3.25042107247001499370e-01,	/* BFD4CD7D, 691CB913 */
146 /* pp2 */	-2.84817495755985104766e-02,	/* BF9D2A51, DBD7194F */
147 /* pp3 */	-5.77027029648944159157e-03,	/* BF77A291, 236668E4 */
148 /* pp4 */	-2.37630166566501626084e-05,	/* BEF8EAD6, 120016AC */
149 /* qq1 */	 3.97917223959155352819e-01,	/* 3FD97779, CDDADC09 */
150 /* qq2 */	 6.50222499887672944485e-02,	/* 3FB0A54C, 5536CEBA */
151 /* qq3 */	 5.08130628187576562776e-03,	/* 3F74D022, C4D36B0F */
152 /* qq4 */	 1.32494738004321644526e-04,	/* 3F215DC9, 221C1A10 */
153 /* qq5 */	-3.96022827877536812320e-06,	/* BED09C43, 42A26120 */
154 /*
155  * Coefficients for approximation to  erf  in [0.84375,1.25]
156  */
157 /* pa0 */	-2.36211856075265944077e-03,	/* BF6359B8, BEF77538 */
158 /* pa1 */	 4.14856118683748331666e-01,	/* 3FDA8D00, AD92B34D */
159 /* pa2 */	-3.72207876035701323847e-01,	/* BFD7D240, FBB8C3F1 */
160 /* pa3 */	 3.18346619901161753674e-01,	/* 3FD45FCA, 805120E4 */
161 /* pa4 */	-1.10894694282396677476e-01,	/* BFBC6398, 3D3E28EC */
162 /* pa5 */	 3.54783043256182359371e-02,	/* 3FA22A36, 599795EB */
163 /* pa6 */	-2.16637559486879084300e-03,	/* BF61BF38, 0A96073F */
164 /* qa1 */	 1.06420880400844228286e-01,	/* 3FBB3E66, 18EEE323 */
165 /* qa2 */	 5.40397917702171048937e-01,	/* 3FE14AF0, 92EB6F33 */
166 /* qa3 */	 7.18286544141962662868e-02,	/* 3FB2635C, D99FE9A7 */
167 /* qa4 */	 1.26171219808761642112e-01,	/* 3FC02660, E763351F */
168 /* qa5 */	 1.36370839120290507362e-02,	/* 3F8BEDC2, 6B51DD1C */
169 /* qa6 */	 1.19844998467991074170e-02,	/* 3F888B54, 5735151D */
170 /*
171  * Coefficients for approximation to  erfc in [1.25,1/0.35]
172  */
173 /* ra0 */	-9.86494403484714822705e-03,	/* BF843412, 600D6435 */
174 /* ra1 */	-6.93858572707181764372e-01,	/* BFE63416, E4BA7360 */
175 /* ra2 */	-1.05586262253232909814e+01,	/* C0251E04, 41B0E726 */
176 /* ra3 */	-6.23753324503260060396e+01,	/* C04F300A, E4CBA38D */
177 /* ra4 */	-1.62396669462573470355e+02,	/* C0644CB1, 84282266 */
178 /* ra5 */	-1.84605092906711035994e+02,	/* C067135C, EBCCABB2 */
179 /* ra6 */	-8.12874355063065934246e+01,	/* C0545265, 57E4D2F2 */
180 /* ra7 */	-9.81432934416914548592e+00,	/* C023A0EF, C69AC25C */
181 /* sa1 */	 1.96512716674392571292e+01,	/* 4033A6B9, BD707687 */
182 /* sa2 */	 1.37657754143519042600e+02,	/* 4061350C, 526AE721 */
183 /* sa3 */	 4.34565877475229228821e+02,	/* 407B290D, D58A1A71 */
184 /* sa4 */	 6.45387271733267880336e+02,	/* 40842B19, 21EC2868 */
185 /* sa5 */	 4.29008140027567833386e+02,	/* 407AD021, 57700314 */
186 /* sa6 */	 1.08635005541779435134e+02,	/* 405B28A3, EE48AE2C */
187 /* sa7 */	 6.57024977031928170135e+00,	/* 401A47EF, 8E484A93 */
188 /* sa8 */	-6.04244152148580987438e-02,	/* BFAEEFF2, EE749A62 */
189 /*
190  * Coefficients for approximation to  erfc in [1/.35,28]
191  */
192 /* rb0 */	-9.86494292470009928597e-03,	/* BF843412, 39E86F4A */
193 /* rb1 */	-7.99283237680523006574e-01,	/* BFE993BA, 70C285DE */
194 /* rb2 */	-1.77579549177547519889e+01,	/* C031C209, 555F995A */
195 /* rb3 */	-1.60636384855821916062e+02,	/* C064145D, 43C5ED98 */
196 /* rb4 */	-6.37566443368389627722e+02,	/* C083EC88, 1375F228 */
197 /* rb5 */	-1.02509513161107724954e+03,	/* C0900461, 6A2E5992 */
198 /* rb6 */	-4.83519191608651397019e+02,	/* C07E384E, 9BDC383F */
199 /* sb1 */	 3.03380607434824582924e+01,	/* 403E568B, 261D5190 */
200 /* sb2 */	 3.25792512996573918826e+02,	/* 40745CAE, 221B9F0A */
201 /* sb3 */	 1.53672958608443695994e+03,	/* 409802EB, 189D5118 */
202 /* sb4 */	 3.19985821950859553908e+03,	/* 40A8FFB7, 688C246A */
203 /* sb5 */	 2.55305040643316442583e+03,	/* 40A3F219, CEDF3BE6 */
204 /* sb6 */	 4.74528541206955367215e+02,	/* 407DA874, E79FE763 */
205 /* sb7 */	-2.24409524465858183362e+01	/* C03670E2, 42712D62 */
206 };
207 
208 #define	tiny	xxx[0]
209 #define	half	xxx[1]
210 #define	one	xxx[2]
211 #define	two	xxx[3]
212 #define	erx	xxx[4]
213 /*
214  * Coefficients for approximation to  erf on [0,0.84375]
215  */
216 #define	efx	xxx[5]
217 #define	efx8	xxx[6]
218 #define	pp0	xxx[7]
219 #define	pp1	xxx[8]
220 #define	pp2	xxx[9]
221 #define	pp3	xxx[10]
222 #define	pp4	xxx[11]
223 #define	qq1	xxx[12]
224 #define	qq2	xxx[13]
225 #define	qq3	xxx[14]
226 #define	qq4	xxx[15]
227 #define	qq5	xxx[16]
228 /*
229  * Coefficients for approximation to  erf  in [0.84375,1.25]
230  */
231 #define	pa0	xxx[17]
232 #define	pa1	xxx[18]
233 #define	pa2	xxx[19]
234 #define	pa3	xxx[20]
235 #define	pa4	xxx[21]
236 #define	pa5	xxx[22]
237 #define	pa6	xxx[23]
238 #define	qa1	xxx[24]
239 #define	qa2	xxx[25]
240 #define	qa3	xxx[26]
241 #define	qa4	xxx[27]
242 #define	qa5	xxx[28]
243 #define	qa6	xxx[29]
244 /*
245  * Coefficients for approximation to  erfc in [1.25,1/0.35]
246  */
247 #define	ra0	xxx[30]
248 #define	ra1	xxx[31]
249 #define	ra2	xxx[32]
250 #define	ra3	xxx[33]
251 #define	ra4	xxx[34]
252 #define	ra5	xxx[35]
253 #define	ra6	xxx[36]
254 #define	ra7	xxx[37]
255 #define	sa1	xxx[38]
256 #define	sa2	xxx[39]
257 #define	sa3	xxx[40]
258 #define	sa4	xxx[41]
259 #define	sa5	xxx[42]
260 #define	sa6	xxx[43]
261 #define	sa7	xxx[44]
262 #define	sa8	xxx[45]
263 /*
264  * Coefficients for approximation to  erfc in [1/.35,28]
265  */
266 #define	rb0	xxx[46]
267 #define	rb1	xxx[47]
268 #define	rb2	xxx[48]
269 #define	rb3	xxx[49]
270 #define	rb4	xxx[50]
271 #define	rb5	xxx[51]
272 #define	rb6	xxx[52]
273 #define	sb1	xxx[53]
274 #define	sb2	xxx[54]
275 #define	sb3	xxx[55]
276 #define	sb4	xxx[56]
277 #define	sb5	xxx[57]
278 #define	sb6	xxx[58]
279 #define	sb7	xxx[59]
280 
281 double
282 erf(double x) {
283 	int hx, ix, i;
284 	double R, S, P, Q, s, y, z, r;
285 
286 	hx = ((int *) &x)[HIWORD];
287 	ix = hx & 0x7fffffff;
288 	if (ix >= 0x7ff00000) {	/* erf(nan)=nan */
289 #if defined(FPADD_TRAPS_INCOMPLETE_ON_NAN)
290 		if (ix >= 0x7ff80000)		/* assumes sparc-like QNaN */
291 			return (x);
292 #endif
293 		i = ((unsigned) hx >> 31) << 1;
294 		return ((double) (1 - i) + one / x);	/* erf(+-inf)=+-1 */
295 	}
296 
297 	if (ix < 0x3feb0000) {	/* |x|<0.84375 */
298 		if (ix < 0x3e300000) {	/* |x|<2**-28 */
299 			if (ix < 0x00800000)	/* avoid underflow */
300 				return (0.125 * (8.0 * x + efx8 * x));
301 			return (x + efx * x);
302 		}
303 		z = x * x;
304 		r = pp0 + z * (pp1 + z * (pp2 + z * (pp3 + z * pp4)));
305 		s = one +
306 			z *(qq1 + z * (qq2 + z * (qq3 + z * (qq4 + z * qq5))));
307 		y = r / s;
308 		return (x + x * y);
309 	}
310 	if (ix < 0x3ff40000) {	/* 0.84375 <= |x| < 1.25 */
311 		s = fabs(x) - one;
312 		P = pa0 + s * (pa1 + s * (pa2 + s * (pa3 + s * (pa4 +
313 			s * (pa5 + s * pa6)))));
314 		Q = one + s * (qa1 + s * (qa2 + s * (qa3 + s * (qa4 +
315 			s * (qa5 + s * qa6)))));
316 		if (hx >= 0)
317 			return (erx + P / Q);
318 		else
319 			return (-erx - P / Q);
320 	}
321 	if (ix >= 0x40180000) {	/* inf > |x| >= 6 */
322 		if (hx >= 0)
323 			return (one - tiny);
324 		else
325 			return (tiny - one);
326 	}
327 	x = fabs(x);
328 	s = one / (x * x);
329 	if (ix < 0x4006DB6E) {	/* |x| < 1/0.35 */
330 		R = ra0 + s * (ra1 + s * (ra2 + s * (ra3 + s * (ra4 +
331 			s * (ra5 + s * (ra6 + s * ra7))))));
332 		S = one + s * (sa1 + s * (sa2 + s * (sa3 + s * (sa4 +
333 			s * (sa5 + s * (sa6 + s * (sa7 + s * sa8)))))));
334 	} else {			/* |x| >= 1/0.35 */
335 		R = rb0 + s * (rb1 + s * (rb2 + s * (rb3 + s * (rb4 +
336 			s * (rb5 + s * rb6)))));
337 		S = one + s * (sb1 + s * (sb2 + s * (sb3 + s * (sb4 +
338 			s * (sb5 + s * (sb6 + s * sb7))))));
339 	}
340 	z = x;
341 	((int *) &z)[LOWORD] = 0;
342 	r = exp(-z * z - 0.5625) * exp((z - x) * (z + x) + R / S);
343 	if (hx >= 0)
344 		return (one - r / x);
345 	else
346 		return (r / x - one);
347 }
348 
349 double
350 erfc(double x) {
351 	int hx, ix;
352 	double R, S, P, Q, s, y, z, r;
353 
354 	hx = ((int *) &x)[HIWORD];
355 	ix = hx & 0x7fffffff;
356 	if (ix >= 0x7ff00000) {	/* erfc(nan)=nan */
357 #if defined(FPADD_TRAPS_INCOMPLETE_ON_NAN)
358 		if (ix >= 0x7ff80000)		/* assumes sparc-like QNaN */
359 			return (x);
360 #endif
361 		/* erfc(+-inf)=0,2 */
362 		return ((double) (((unsigned) hx >> 31) << 1) + one / x);
363 	}
364 
365 	if (ix < 0x3feb0000) {	/* |x| < 0.84375 */
366 		if (ix < 0x3c700000)	/* |x| < 2**-56 */
367 			return (one - x);
368 		z = x * x;
369 		r = pp0 + z * (pp1 + z * (pp2 + z * (pp3 + z * pp4)));
370 		s = one +
371 			z * (qq1 + z * (qq2 + z * (qq3 + z * (qq4 + z * qq5))));
372 		y = r / s;
373 		if (hx < 0x3fd00000) {	/* x < 1/4 */
374 			return (one - (x + x * y));
375 		} else {
376 			r = x * y;
377 			r += (x - half);
378 			return (half - r);
379 		}
380 	}
381 	if (ix < 0x3ff40000) {	/* 0.84375 <= |x| < 1.25 */
382 		s = fabs(x) - one;
383 		P = pa0 + s * (pa1 + s * (pa2 + s * (pa3 + s * (pa4 +
384 			s * (pa5 + s * pa6)))));
385 		Q = one + s * (qa1 + s * (qa2 + s * (qa3 + s * (qa4 +
386 			s * (qa5 + s * qa6)))));
387 		if (hx >= 0) {
388 			z = one - erx;
389 			return (z - P / Q);
390 		} else {
391 			z = erx + P / Q;
392 			return (one + z);
393 		}
394 	}
395 	if (ix < 0x403c0000) {	/* |x|<28 */
396 		x = fabs(x);
397 		s = one / (x * x);
398 		if (ix < 0x4006DB6D) {	/* |x| < 1/.35 ~ 2.857143 */
399 			R = ra0 + s * (ra1 + s * (ra2 + s * (ra3 + s * (ra4 +
400 				s * (ra5 + s * (ra6 + s * ra7))))));
401 			S = one + s * (sa1 + s * (sa2 + s * (sa3 + s * (sa4 +
402 				s * (sa5 + s * (sa6 + s * (sa7 + s * sa8)))))));
403 		} else {
404 			/* |x| >= 1/.35 ~ 2.857143 */
405 			if (hx < 0 && ix >= 0x40180000)
406 				return (two - tiny);	/* x < -6 */
407 
408 			R = rb0 + s * (rb1 + s * (rb2 + s * (rb3 + s * (rb4 +
409 				s * (rb5 + s * rb6)))));
410 			S = one + s * (sb1 + s * (sb2 + s * (sb3 + s * (sb4 +
411 				s * (sb5 + s * (sb6 + s * sb7))))));
412 		}
413 		z = x;
414 		((int *) &z)[LOWORD] = 0;
415 		r = exp(-z * z - 0.5625) * exp((z - x) * (z + x) + R / S);
416 		if (hx > 0)
417 			return (r / x);
418 		else
419 			return (two - r / x);
420 	} else {
421 		if (hx > 0)
422 			return (tiny * tiny);
423 		else
424 			return (two - tiny);
425 	}
426 }
427