s_erf.c revision 97409
18478Swollman/* @(#)s_erf.c 5.1 93/09/24 */
28478Swollman/*
38478Swollman * ====================================================
48478Swollman * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
58478Swollman *
68478Swollman * Developed at SunPro, a Sun Microsystems, Inc. business.
78478Swollman * Permission to use, copy, modify, and distribute this
88478Swollman * software is freely granted, provided that this notice
98478Swollman * is preserved.
108478Swollman * ====================================================
118478Swollman */
128478Swollman
138478Swollman#ifndef lint
148478Swollmanstatic char rcsid[] = "$FreeBSD: head/lib/msun/src/s_erf.c 97409 2002-05-28 17:51:46Z alfred $";
158478Swollman#endif
168478Swollman
178478Swollman/* double erf(double x)
188478Swollman * double erfc(double x)
198478Swollman *			     x
208478Swollman *		      2      |\
218478Swollman *     erf(x)  =  ---------  | exp(-t*t)dt
228478Swollman *	 	   sqrt(pi) \|
238478Swollman *			     0
248478Swollman *
258478Swollman *     erfc(x) =  1-erf(x)
268478Swollman *  Note that
278478Swollman *		erf(-x) = -erf(x)
288478Swollman *		erfc(-x) = 2 - erfc(x)
298478Swollman *
30114589Sobrien * Method:
318478Swollman *	1. For |x| in [0, 0.84375]
3236999Scharnier *	    erf(x)  = x + x*R(x^2)
338478Swollman *          erfc(x) = 1 - erf(x)           if x in [-.84375,0.25]
348478Swollman *                  = 0.5 + ((0.5-x)-x*R)  if x in [0.25,0.84375]
358478Swollman *	   where R = P/Q where P is an odd poly of degree 8 and
368478Swollman *	   Q is an odd poly of degree 10.
378478Swollman *						 -57.90
3836999Scharnier *			| R - (erf(x)-x)/x | <= 2
39114589Sobrien *
4036999Scharnier *
41114589Sobrien *	   Remark. The formula is derived by noting
42114589Sobrien *          erf(x) = (2/sqrt(pi))*(x - x^3/3 + x^5/10 - x^7/42 + ....)
438478Swollman *	   and that
44136104Sdes *          2/sqrt(pi) = 1.128379167095512573896158903121545171688
45136104Sdes *	   is close to one. The interval is chosen because the fix
46136104Sdes *	   point of erf(x) is near 0.6174 (i.e., erf(x)=x when x is
47136104Sdes *	   near 0.6174), and by some experiment, 0.84375 is chosen to
4836999Scharnier * 	   guarantee the error is less than one ulp for erf.
49242451Salfred *
50136104Sdes *      2. For |x| in [0.84375,1.25], let s = |x| - 1, and
51136104Sdes *         c = 0.84506291151 rounded to single (24 bits)
52136104Sdes *         	erf(x)  = sign(x) * (c  + P1(s)/Q1(s))
538478Swollman *         	erfc(x) = (1-c)  - P1(s)/Q1(s) if x > 0
5478732Sdd *			  1+(c+P1(s)/Q1(s))    if x < 0
5596381Salfred *         	|P1/Q1 - (erf(|x|)-c)| <= 2**-59.06
56136104Sdes *	   Remark: here we use the taylor series expansion at x=1.
578478Swollman *		erf(1+s) = erf(1) + s*Poly(s)
588478Swollman *			 = 0.845.. + P1(s)/Q1(s)
59136104Sdes *	   That is, we use rational approximation to approximate
608478Swollman *			erf(1+s) - (c = (single)0.84506291151)
61136104Sdes *	   Note that |P1/Q1|< 0.078 for x in [0.84375,1.25]
62136104Sdes *	   where
63136104Sdes *		P1(s) = degree 6 poly in s
64242451Salfred *		Q1(s) = degree 6 poly in s
65136104Sdes *
66242451Salfred *      3. For x in [1.25,1/0.35(~2.857143)],
67242486Salfred *         	erfc(x) = (1/x)*exp(-x*x-0.5625+R1/S1)
68136104Sdes *         	erf(x)  = 1 - erfc(x)
69136104Sdes *	   where
70136104Sdes *		R1(z) = degree 7 poly in z, (z=1/x^2)
71136104Sdes *		S1(z) = degree 8 poly in z
72136104Sdes *
73136104Sdes *      4. For x in [1/0.35,28]
74136104Sdes *         	erfc(x) = (1/x)*exp(-x*x-0.5625+R2/S2) if x > 0
75163852Sjhb *			= 2.0 - (1/x)*exp(-x*x-0.5625+R2/S2) if -6<x<0
76136104Sdes *			= 2.0 - tiny		(if x <= -6)
77158083Sps *         	erf(x)  = sign(x)*(1.0 - erfc(x)) if x < 6, else
78136104Sdes *         	erf(x)  = sign(x)*(1.0 - tiny)
79158083Sps *	   where
80136104Sdes *		R2(z) = degree 6 poly in z, (z=1/x^2)
81163852Sjhb *		S2(z) = degree 7 poly in z
82158083Sps *
83158083Sps *      Note1:
84158083Sps *	   To compute exp(-x*x-0.5625+R/S), let s be a single
85163852Sjhb *	   precision number and s := x; then
86136104Sdes *		-x*x = -s*s + (s-x)*(s+x)
87136104Sdes *	        exp(-x*x-0.5626+R/S) =
88136104Sdes *			exp(-s*s-0.5625)*exp((s-x)*(s+x)+R/S);
89136104Sdes *      Note2:
90136110Sdes *	   Here 4 and 5 make use of the asymptotic series
91136104Sdes *			  exp(-x*x)
92136104Sdes *		erfc(x) ~ ---------- * ( 1 + Poly(1/x^2) )
93136104Sdes *			  x*sqrt(pi)
94136104Sdes *	   We use rational approximation to approximate
95136104Sdes *      	g(s)=f(1/x^2) = log(erfc(x)*x) - x*x + 0.5625
96136104Sdes *	   Here is the error bound for R1/S1 and R2/S2
97242451Salfred *      	|R1/S1 - f(x)|  < 2**(-62.57)
98242451Salfred *      	|R2/S2 - f(x)|  < 2**(-61.52)
99242451Salfred *
100242451Salfred *      5. For inf > x >= 28
101242451Salfred *         	erf(x)  = sign(x) *(1 - tiny)  (raise inexact)
102242451Salfred *         	erfc(x) = tiny*tiny (raise underflow) if x > 0
103242451Salfred *			= 2 - tiny if x<0
104242451Salfred *
105242451Salfred *      7. Special case:
106242451Salfred *         	erf(0)  = 0, erf(inf)  = 1, erf(-inf) = -1,
107242451Salfred *         	erfc(0) = 1, erfc(inf) = 0, erfc(-inf) = 2,
108242451Salfred *	   	erfc/erf(NaN) is NaN
109242451Salfred */
110242451Salfred
111242451Salfred
112242451Salfred#include "math.h"
113242451Salfred#include "math_private.h"
114242451Salfred
115242451Salfredstatic const double
116242451Salfredtiny	    = 1e-300,
117242451Salfredhalf=  5.00000000000000000000e-01, /* 0x3FE00000, 0x00000000 */
118242451Salfredone =  1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */
119242451Salfredtwo =  2.00000000000000000000e+00, /* 0x40000000, 0x00000000 */
120242451Salfred	/* c = (float)0.84506291151 */
121242451Salfrederx =  8.45062911510467529297e-01, /* 0x3FEB0AC1, 0x60000000 */
122242451Salfred/*
1238478Swollman * Coefficients for approximation to  erf on [0,0.84375]
12492542Simp */
1258478Swollmanefx =  1.28379167095512586316e-01, /* 0x3FC06EBA, 0x8214DB69 */
126136104Sdesefx8=  1.02703333676410069053e+00, /* 0x3FF06EBA, 0x8214DB69 */
12793491Sphkpp0  =  1.28379167095512558561e-01, /* 0x3FC06EBA, 0x8214DB68 */
12893491Sphkpp1  = -3.25042107247001499370e-01, /* 0xBFD4CD7D, 0x691CB913 */
129242451Salfredpp2  = -2.84817495755985104766e-02, /* 0xBF9D2A51, 0xDBD7194F */
1308478Swollmanpp3  = -5.77027029648944159157e-03, /* 0xBF77A291, 0x236668E4 */
131242451Salfredpp4  = -2.37630166566501626084e-05, /* 0xBEF8EAD6, 0x120016AC */
1328478Swollmanqq1  =  3.97917223959155352819e-01, /* 0x3FD97779, 0xCDDADC09 */
133242451Salfredqq2  =  6.50222499887672944485e-02, /* 0x3FB0A54C, 0x5536CEBA */
134242451Salfredqq3  =  5.08130628187576562776e-03, /* 0x3F74D022, 0xC4D36B0F */
135242451Salfredqq4  =  1.32494738004321644526e-04, /* 0x3F215DC9, 0x221C1A10 */
1368478Swollmanqq5  = -3.96022827877536812320e-06, /* 0xBED09C43, 0x42A26120 */
1378478Swollman/*
1388478Swollman * Coefficients for approximation to  erf  in [0.84375,1.25]
1398478Swollman */
1408478Swollmanpa0  = -2.36211856075265944077e-03, /* 0xBF6359B8, 0xBEF77538 */
1418478Swollmanpa1  =  4.14856118683748331666e-01, /* 0x3FDA8D00, 0xAD92B34D */
142136104Sdespa2  = -3.72207876035701323847e-01, /* 0xBFD7D240, 0xFBB8C3F1 */
143136104Sdespa3  =  3.18346619901161753674e-01, /* 0x3FD45FCA, 0x805120E4 */
1448478Swollmanpa4  = -1.10894694282396677476e-01, /* 0xBFBC6398, 0x3D3E28EC */
1458478Swollmanpa5  =  3.54783043256182359371e-02, /* 0x3FA22A36, 0x599795EB */
146242451Salfredpa6  = -2.16637559486879084300e-03, /* 0xBF61BF38, 0x0A96073F */
147242451Salfredqa1  =  1.06420880400844228286e-01, /* 0x3FBB3E66, 0x18EEE323 */
148242451Salfredqa2  =  5.40397917702171048937e-01, /* 0x3FE14AF0, 0x92EB6F33 */
149242451Salfredqa3  =  7.18286544141962662868e-02, /* 0x3FB2635C, 0xD99FE9A7 */
150242451Salfredqa4  =  1.26171219808761642112e-01, /* 0x3FC02660, 0xE763351F */
151136104Sdesqa5  =  1.36370839120290507362e-02, /* 0x3F8BEDC2, 0x6B51DD1C */
1528478Swollmanqa6  =  1.19844998467991074170e-02, /* 0x3F888B54, 0x5735151D */
1538478Swollman/*
154136104Sdes * Coefficients for approximation to  erfc in [1.25,1/0.35]
155291480Ssmh */
156291480Ssmhra0  = -9.86494403484714822705e-03, /* 0xBF843412, 0x600D6435 */
157291480Ssmhra1  = -6.93858572707181764372e-01, /* 0xBFE63416, 0xE4BA7360 */
158291480Ssmhra2  = -1.05586262253232909814e+01, /* 0xC0251E04, 0x41B0E726 */
159291480Ssmhra3  = -6.23753324503260060396e+01, /* 0xC04F300A, 0xE4CBA38D */
160291480Ssmhra4  = -1.62396669462573470355e+02, /* 0xC0644CB1, 0x84282266 */
161291480Ssmhra5  = -1.84605092906711035994e+02, /* 0xC067135C, 0xEBCCABB2 */
162291480Ssmhra6  = -8.12874355063065934246e+01, /* 0xC0545265, 0x57E4D2F2 */
163291480Ssmhra7  = -9.81432934416914548592e+00, /* 0xC023A0EF, 0xC69AC25C */
164291480Ssmhsa1  =  1.96512716674392571292e+01, /* 0x4033A6B9, 0xBD707687 */
165291480Ssmhsa2  =  1.37657754143519042600e+02, /* 0x4061350C, 0x526AE721 */
166291480Ssmhsa3  =  4.34565877475229228821e+02, /* 0x407B290D, 0xD58A1A71 */
167291480Ssmhsa4  =  6.45387271733267880336e+02, /* 0x40842B19, 0x21EC2868 */
168291480Ssmhsa5  =  4.29008140027567833386e+02, /* 0x407AD021, 0x57700314 */
169291480Ssmhsa6  =  1.08635005541779435134e+02, /* 0x405B28A3, 0xEE48AE2C */
170291480Ssmhsa7  =  6.57024977031928170135e+00, /* 0x401A47EF, 0x8E484A93 */
17193491Sphksa8  = -6.04244152148580987438e-02, /* 0xBFAEEFF2, 0xEE749A62 */
172291480Ssmh/*
173291480Ssmh * Coefficients for approximation to  erfc in [1/.35,28]
17493491Sphk */
17594272Sphkrb0  = -9.86494292470009928597e-03, /* 0xBF843412, 0x39E86F4A */
17693491Sphkrb1  = -7.99283237680523006574e-01, /* 0xBFE993BA, 0x70C285DE */
17794272Sphkrb2  = -1.77579549177547519889e+01, /* 0xC031C209, 0x555F995A */
17893491Sphkrb3  = -1.60636384855821916062e+02, /* 0xC064145D, 0x43C5ED98 */
179291480Ssmhrb4  = -6.37566443368389627722e+02, /* 0xC083EC88, 0x1375F228 */
1808478Swollmanrb5  = -1.02509513161107724954e+03, /* 0xC0900461, 0x6A2E5992 */
18193491Sphkrb6  = -4.83519191608651397019e+02, /* 0xC07E384E, 0x9BDC383F */
18293491Sphksb1  =  3.03380607434824582924e+01, /* 0x403E568B, 0x261D5190 */
18393491Sphksb2  =  3.25792512996573918826e+02, /* 0x40745CAE, 0x221B9F0A */
18493491Sphksb3  =  1.53672958608443695994e+03, /* 0x409802EB, 0x189D5118 */
18594272Sphksb4  =  3.19985821950859553908e+03, /* 0x40A8FFB7, 0x688C246A */
18693491Sphksb5  =  2.55305040643316442583e+03, /* 0x40A3F219, 0xCEDF3BE6 */
18793491Sphksb6  =  4.74528541206955367215e+02, /* 0x407DA874, 0xE79FE763 */
1888478Swollmansb7  = -2.24409524465858183362e+01; /* 0xC03670E2, 0x42712D62 */
18993491Sphk
19094272Sphk	double erf(double x)
1918478Swollman{
19293491Sphk	int32_t hx,ix,i;
1938478Swollman	double R,S,P,Q,s,y,z,r;
194	GET_HIGH_WORD(hx,x);
195	ix = hx&0x7fffffff;
196	if(ix>=0x7ff00000) {		/* erf(nan)=nan */
197	    i = ((u_int32_t)hx>>31)<<1;
198	    return (double)(1-i)+one/x;	/* erf(+-inf)=+-1 */
199	}
200
201	if(ix < 0x3feb0000) {		/* |x|<0.84375 */
202	    if(ix < 0x3e300000) { 	/* |x|<2**-28 */
203	        if (ix < 0x00800000)
204		    return 0.125*(8.0*x+efx8*x);  /*avoid underflow */
205		return x + efx*x;
206	    }
207	    z = x*x;
208	    r = pp0+z*(pp1+z*(pp2+z*(pp3+z*pp4)));
209	    s = one+z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5))));
210	    y = r/s;
211	    return x + x*y;
212	}
213	if(ix < 0x3ff40000) {		/* 0.84375 <= |x| < 1.25 */
214	    s = fabs(x)-one;
215	    P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))));
216	    Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))));
217	    if(hx>=0) return erx + P/Q; else return -erx - P/Q;
218	}
219	if (ix >= 0x40180000) {		/* inf>|x|>=6 */
220	    if(hx>=0) return one-tiny; else return tiny-one;
221	}
222	x = fabs(x);
223 	s = one/(x*x);
224	if(ix< 0x4006DB6E) {	/* |x| < 1/0.35 */
225	    R=ra0+s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*(
226				ra5+s*(ra6+s*ra7))))));
227	    S=one+s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*(
228				sa5+s*(sa6+s*(sa7+s*sa8)))))));
229	} else {	/* |x| >= 1/0.35 */
230	    R=rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(
231				rb5+s*rb6)))));
232	    S=one+s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*(
233				sb5+s*(sb6+s*sb7))))));
234	}
235	z  = x;
236	SET_LOW_WORD(z,0);
237	r  =  __ieee754_exp(-z*z-0.5625)*__ieee754_exp((z-x)*(z+x)+R/S);
238	if(hx>=0) return one-r/x; else return  r/x-one;
239}
240
241	double erfc(double x)
242{
243	int32_t hx,ix;
244	double R,S,P,Q,s,y,z,r;
245	GET_HIGH_WORD(hx,x);
246	ix = hx&0x7fffffff;
247	if(ix>=0x7ff00000) {			/* erfc(nan)=nan */
248						/* erfc(+-inf)=0,2 */
249	    return (double)(((u_int32_t)hx>>31)<<1)+one/x;
250	}
251
252	if(ix < 0x3feb0000) {		/* |x|<0.84375 */
253	    if(ix < 0x3c700000)  	/* |x|<2**-56 */
254		return one-x;
255	    z = x*x;
256	    r = pp0+z*(pp1+z*(pp2+z*(pp3+z*pp4)));
257	    s = one+z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5))));
258	    y = r/s;
259	    if(hx < 0x3fd00000) {  	/* x<1/4 */
260		return one-(x+x*y);
261	    } else {
262		r = x*y;
263		r += (x-half);
264	        return half - r ;
265	    }
266	}
267	if(ix < 0x3ff40000) {		/* 0.84375 <= |x| < 1.25 */
268	    s = fabs(x)-one;
269	    P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))));
270	    Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))));
271	    if(hx>=0) {
272	        z  = one-erx; return z - P/Q;
273	    } else {
274		z = erx+P/Q; return one+z;
275	    }
276	}
277	if (ix < 0x403c0000) {		/* |x|<28 */
278	    x = fabs(x);
279 	    s = one/(x*x);
280	    if(ix< 0x4006DB6D) {	/* |x| < 1/.35 ~ 2.857143*/
281	        R=ra0+s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*(
282				ra5+s*(ra6+s*ra7))))));
283	        S=one+s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*(
284				sa5+s*(sa6+s*(sa7+s*sa8)))))));
285	    } else {			/* |x| >= 1/.35 ~ 2.857143 */
286		if(hx<0&&ix>=0x40180000) return two-tiny;/* x < -6 */
287	        R=rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(
288				rb5+s*rb6)))));
289	        S=one+s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*(
290				sb5+s*(sb6+s*sb7))))));
291	    }
292	    z  = x;
293	    SET_LOW_WORD(z,0);
294	    r  =  __ieee754_exp(-z*z-0.5625)*
295			__ieee754_exp((z-x)*(z+x)+R/S);
296	    if(hx>0) return r/x; else return two-r/x;
297	} else {
298	    if(hx>0) return tiny*tiny; else return two-tiny;
299	}
300}
301