s_erf.c revision 268588
177701Sbrian/* @(#)s_erf.c 5.1 93/09/24 */
285964Sbrian/*
377701Sbrian * ====================================================
477701Sbrian * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
577701Sbrian *
677701Sbrian * Developed at SunPro, a Sun Microsystems, Inc. business.
777701Sbrian * Permission to use, copy, modify, and distribute this
877701Sbrian * software is freely granted, provided that this notice
977701Sbrian * is preserved.
1077701Sbrian * ====================================================
1177701Sbrian */
1277701Sbrian
1377701Sbrian#include <sys/cdefs.h>
1477701Sbrian__FBSDID("$FreeBSD: head/lib/msun/src/s_erf.c 268588 2014-07-13 15:45:45Z kargl $");
1577701Sbrian
1677701Sbrian/* double erf(double x)
1777701Sbrian * double erfc(double x)
1877701Sbrian *			     x
1977701Sbrian *		      2      |\
2077701Sbrian *     erf(x)  =  ---------  | exp(-t*t)dt
2177701Sbrian *	 	   sqrt(pi) \|
2277701Sbrian *			     0
2377701Sbrian *
2477701Sbrian *     erfc(x) =  1-erf(x)
2577701Sbrian *  Note that
2677701Sbrian *		erf(-x) = -erf(x)
2784195Sdillon *		erfc(-x) = 2 - erfc(x)
2884195Sdillon *
2984195Sdillon * Method:
3026026Sbrian *	1. For |x| in [0, 0.84375]
3126026Sbrian *	    erf(x)  = x + x*R(x^2)
3226026Sbrian *          erfc(x) = 1 - erf(x)           if x in [-.84375,0.25]
3326026Sbrian *                  = 0.5 + ((0.5-x)-x*R)  if x in [0.25,0.84375]
3426026Sbrian *	   where R = P/Q where P is an odd poly of degree 8 and
3526026Sbrian *	   Q is an odd poly of degree 10.
3626026Sbrian *						 -57.90
3726026Sbrian *			| R - (erf(x)-x)/x | <= 2
3826026Sbrian *
3926026Sbrian *
4026026Sbrian *	   Remark. The formula is derived by noting
4126026Sbrian *          erf(x) = (2/sqrt(pi))*(x - x^3/3 + x^5/10 - x^7/42 + ....)
4226026Sbrian *	   and that
4326026Sbrian *          2/sqrt(pi) = 1.128379167095512573896158903121545171688
4459047Sru *	   is close to one. The interval is chosen because the fix
4526026Sbrian *	   point of erf(x) is near 0.6174 (i.e., erf(x)=x when x is
4626026Sbrian *	   near 0.6174), and by some experiment, 0.84375 is chosen to
4726026Sbrian * 	   guarantee the error is less than one ulp for erf.
4826026Sbrian *
4926026Sbrian *      2. For |x| in [0.84375,1.25], let s = |x| - 1, and
50131612Sdes *         c = 0.84506291151 rounded to single (24 bits)
51131612Sdes *         	erf(x)  = sign(x) * (c  + P1(s)/Q1(s))
52131612Sdes *         	erfc(x) = (1-c)  - P1(s)/Q1(s) if x > 0
5326026Sbrian *			  1+(c+P1(s)/Q1(s))    if x < 0
5426026Sbrian *         	|P1/Q1 - (erf(|x|)-c)| <= 2**-59.06
55131612Sdes *	   Remark: here we use the taylor series expansion at x=1.
56131612Sdes *		erf(1+s) = erf(1) + s*Poly(s)
5726026Sbrian *			 = 0.845.. + P1(s)/Q1(s)
5826026Sbrian *	   That is, we use rational approximation to approximate
59131612Sdes *			erf(1+s) - (c = (single)0.84506291151)
60131612Sdes *	   Note that |P1/Q1|< 0.078 for x in [0.84375,1.25]
61131612Sdes *	   where
62131612Sdes *		P1(s) = degree 6 poly in s
63131612Sdes *		Q1(s) = degree 6 poly in s
64131612Sdes *
65131612Sdes *      3. For x in [1.25,1/0.35(~2.857143)],
66131612Sdes *         	erfc(x) = (1/x)*exp(-x*x-0.5625+R1/S1)
6726026Sbrian *         	erf(x)  = 1 - erfc(x)
6826026Sbrian *	   where
69131612Sdes *		R1(z) = degree 7 poly in z, (z=1/x^2)
70131612Sdes *		S1(z) = degree 8 poly in z
7126026Sbrian *
7226026Sbrian *      4. For x in [1/0.35,28]
73131612Sdes *         	erfc(x) = (1/x)*exp(-x*x-0.5625+R2/S2) if x > 0
74131612Sdes *			= 2.0 - (1/x)*exp(-x*x-0.5625+R2/S2) if -6<x<0
7526026Sbrian *			= 2.0 - tiny		(if x <= -6)
7626026Sbrian *         	erf(x)  = sign(x)*(1.0 - erfc(x)) if x < 6, else
77131612Sdes *         	erf(x)  = sign(x)*(1.0 - tiny)
78131612Sdes *	   where
79131612Sdes *		R2(z) = degree 6 poly in z, (z=1/x^2)
80131612Sdes *		S2(z) = degree 7 poly in z
81131612Sdes *
82131612Sdes *      Note1:
83131612Sdes *	   To compute exp(-x*x-0.5625+R/S), let s be a single
84131612Sdes *	   precision number and s := x; then
85131612Sdes *		-x*x = -s*s + (s-x)*(s+x)
8626026Sbrian *	        exp(-x*x-0.5626+R/S) =
8726026Sbrian *			exp(-s*s-0.5625)*exp((s-x)*(s+x)+R/S);
88131612Sdes *      Note2:
89131612Sdes *	   Here 4 and 5 make use of the asymptotic series
90131612Sdes *			  exp(-x*x)
91131612Sdes *		erfc(x) ~ ---------- * ( 1 + Poly(1/x^2) )
92131612Sdes *			  x*sqrt(pi)
93131612Sdes *	   We use rational approximation to approximate
9427864Sbrian *      	g(s)=f(1/x^2) = log(erfc(x)*x) - x*x + 0.5625
9527864Sbrian *	   Here is the error bound for R1/S1 and R2/S2
96131612Sdes *      	|R1/S1 - f(x)|  < 2**(-62.57)
97131612Sdes *      	|R2/S2 - f(x)|  < 2**(-61.52)
98131612Sdes *
99131612Sdes *      5. For inf > x >= 28
10027864Sbrian *         	erf(x)  = sign(x) *(1 - tiny)  (raise inexact)
10141759Sdillon *         	erfc(x) = tiny*tiny (raise underflow) if x > 0
10241759Sdillon *			= 2 - tiny if x<0
10341759Sdillon *
10499207Sbrian *      7. Special case:
10561861Sru *         	erf(0)  = 0, erf(inf)  = 1, erf(-inf) = -1,
10661861Sru *         	erfc(0) = 1, erfc(inf) = 0, erfc(-inf) = 2,
10763899Sarchie *	   	erfc/erf(NaN) is NaN
10863899Sarchie */
10963899Sarchie
11063899Sarchie
11132377Seivind#include "math.h"
11226026Sbrian#include "math_private.h"
11326026Sbrian
114145921Sglebius/* XXX Prevent compilers from erroneously constant folding: */
115145921Sglebiusstatic const volatile double tiny= 1e-300;
116164798Spiso
117164798Spisostatic const double
118188294Spisohalf= 0.5,
119145921Sglebiusone = 1,
12026026Sbriantwo = 2,
121162674Spiso/* c = (float)0.84506291151 */
122145921Sglebiuserx =  8.45062911510467529297e-01, /* 0x3FEB0AC1, 0x60000000 */
123164798Spiso/*
124162674Spiso * In the domain [0, 2**-28], only the first term in the power series
125162674Spiso * expansion of erf(x) is used.  The magnitude of the first neglected
126162674Spiso * terms is less than 2**-84.
127145921Sglebius */
12826026Sbrianefx =  1.28379167095512586316e-01, /* 0x3FC06EBA, 0x8214DB69 */
12926026Sbrianefx8=  1.02703333676410069053e+00, /* 0x3FF06EBA, 0x8214DB69 */
13026026Sbrian/*
13126026Sbrian * Coefficients for approximation to erf on [0,0.84375]
13226026Sbrian */
13326026Sbrianpp0  =  1.28379167095512558561e-01, /* 0x3FC06EBA, 0x8214DB68 */
13426026Sbrianpp1  = -3.25042107247001499370e-01, /* 0xBFD4CD7D, 0x691CB913 */
13526026Sbrianpp2  = -2.84817495755985104766e-02, /* 0xBF9D2A51, 0xDBD7194F */
136145921Sglebiuspp3  = -5.77027029648944159157e-03, /* 0xBF77A291, 0x236668E4 */
137145932Sglebiuspp4  = -2.37630166566501626084e-05, /* 0xBEF8EAD6, 0x120016AC */
138145921Sglebiusqq1  =  3.97917223959155352819e-01, /* 0x3FD97779, 0xCDDADC09 */
139162674Spisoqq2  =  6.50222499887672944485e-02, /* 0x3FB0A54C, 0x5536CEBA */
140145921Sglebiusqq3  =  5.08130628187576562776e-03, /* 0x3F74D022, 0xC4D36B0F */
141162674Spisoqq4  =  1.32494738004321644526e-04, /* 0x3F215DC9, 0x221C1A10 */
142145932Sglebiusqq5  = -3.96022827877536812320e-06, /* 0xBED09C43, 0x42A26120 */
14326026Sbrian/*
144162674Spiso * Coefficients for approximation to erf in [0.84375,1.25]
145145921Sglebius */
14626026Sbrianpa0  = -2.36211856075265944077e-03, /* 0xBF6359B8, 0xBEF77538 */
147188294Spisopa1  =  4.14856118683748331666e-01, /* 0x3FDA8D00, 0xAD92B34D */
148188294Spisopa2  = -3.72207876035701323847e-01, /* 0xBFD7D240, 0xFBB8C3F1 */
149188294Spisopa3  =  3.18346619901161753674e-01, /* 0x3FD45FCA, 0x805120E4 */
150188294Spisopa4  = -1.10894694282396677476e-01, /* 0xBFBC6398, 0x3D3E28EC */
151188294Spisopa5  =  3.54783043256182359371e-02, /* 0x3FA22A36, 0x599795EB */
152188294Spisopa6  = -2.16637559486879084300e-03, /* 0xBF61BF38, 0x0A96073F */
153188294Spisoqa1  =  1.06420880400844228286e-01, /* 0x3FBB3E66, 0x18EEE323 */
154188294Spisoqa2  =  5.40397917702171048937e-01, /* 0x3FE14AF0, 0x92EB6F33 */
155188294Spisoqa3  =  7.18286544141962662868e-02, /* 0x3FB2635C, 0xD99FE9A7 */
156188294Spisoqa4  =  1.26171219808761642112e-01, /* 0x3FC02660, 0xE763351F */
157188294Spisoqa5  =  1.36370839120290507362e-02, /* 0x3F8BEDC2, 0x6B51DD1C */
158127689Sdesqa6  =  1.19844998467991074170e-02, /* 0x3F888B54, 0x5735151D */
159127689Sdes/*
160127689Sdes * Coefficients for approximation to erfc in [1.25,1/0.35]
161131693Sdes */
16226026Sbrianra0  = -9.86494403484714822705e-03, /* 0xBF843412, 0x600D6435 */
163131693Sdesra1  = -6.93858572707181764372e-01, /* 0xBFE63416, 0xE4BA7360 */
164131693Sdesra2  = -1.05586262253232909814e+01, /* 0xC0251E04, 0x41B0E726 */
165131693Sdesra3  = -6.23753324503260060396e+01, /* 0xC04F300A, 0xE4CBA38D */
166131693Sdesra4  = -1.62396669462573470355e+02, /* 0xC0644CB1, 0x84282266 */
167131693Sdesra5  = -1.84605092906711035994e+02, /* 0xC067135C, 0xEBCCABB2 */
168131693Sdesra6  = -8.12874355063065934246e+01, /* 0xC0545265, 0x57E4D2F2 */
169131693Sdesra7  = -9.81432934416914548592e+00, /* 0xC023A0EF, 0xC69AC25C */
170131693Sdessa1  =  1.96512716674392571292e+01, /* 0x4033A6B9, 0xBD707687 */
171127689Sdessa2  =  1.37657754143519042600e+02, /* 0x4061350C, 0x526AE721 */
17226026Sbriansa3  =  4.34565877475229228821e+02, /* 0x407B290D, 0xD58A1A71 */
17326026Sbriansa4  =  6.45387271733267880336e+02, /* 0x40842B19, 0x21EC2868 */
17426026Sbriansa5  =  4.29008140027567833386e+02, /* 0x407AD021, 0x57700314 */
17526026Sbriansa6  =  1.08635005541779435134e+02, /* 0x405B28A3, 0xEE48AE2C */
17644307Sbriansa7  =  6.57024977031928170135e+00, /* 0x401A47EF, 0x8E484A93 */
17726026Sbriansa8  = -6.04244152148580987438e-02, /* 0xBFAEEFF2, 0xEE749A62 */
17851125Sru/*
17926026Sbrian * Coefficients for approximation to erfc in [1/.35,28]
18026026Sbrian */
18126026Sbrianrb0  = -9.86494292470009928597e-03, /* 0xBF843412, 0x39E86F4A */
18226026Sbrianrb1  = -7.99283237680523006574e-01, /* 0xBFE993BA, 0x70C285DE */
18326026Sbrianrb2  = -1.77579549177547519889e+01, /* 0xC031C209, 0x555F995A */
18426026Sbrianrb3  = -1.60636384855821916062e+02, /* 0xC064145D, 0x43C5ED98 */
185176884Spisorb4  = -6.37566443368389627722e+02, /* 0xC083EC88, 0x1375F228 */
18626026Sbrianrb5  = -1.02509513161107724954e+03, /* 0xC0900461, 0x6A2E5992 */
187176884Spisorb6  = -4.83519191608651397019e+02, /* 0xC07E384E, 0x9BDC383F */
18826026Sbriansb1  =  3.03380607434824582924e+01, /* 0x403E568B, 0x261D5190 */
18926026Sbriansb2  =  3.25792512996573918826e+02, /* 0x40745CAE, 0x221B9F0A */
19026026Sbriansb3  =  1.53672958608443695994e+03, /* 0x409802EB, 0x189D5118 */
191176884Spisosb4  =  3.19985821950859553908e+03, /* 0x40A8FFB7, 0x688C246A */
19226026Sbriansb5  =  2.55305040643316442583e+03, /* 0x40A3F219, 0xCEDF3BE6 */
19326026Sbriansb6  =  4.74528541206955367215e+02, /* 0x407DA874, 0xE79FE763 */
194131614Sdessb7  = -2.24409524465858183362e+01; /* 0xC03670E2, 0x42712D62 */
195127094Sdes
196176884Spisodouble
197131614Sdeserf(double x)
198176884Spiso{
199131614Sdes	int32_t hx,ix,i;
200127094Sdes	double R,S,P,Q,s,y,z,r;
201127094Sdes	GET_HIGH_WORD(hx,x);
202176884Spiso	ix = hx&0x7fffffff;
203131614Sdes	if(ix>=0x7ff00000) {		/* erf(nan)=nan */
204127094Sdes	    i = ((u_int32_t)hx>>31)<<1;
205127094Sdes	    return (double)(1-i)+one/x;	/* erf(+-inf)=+-1 */
20626026Sbrian	}
20726026Sbrian
20826026Sbrian	if(ix < 0x3feb0000) {		/* |x|<0.84375 */
209176884Spiso	    if(ix < 0x3e300000) { 	/* |x|<2**-28 */
21026026Sbrian	        if (ix < 0x00800000)
21126026Sbrian		    return (8*x+efx8*x)/8;	/* avoid spurious underflow */
212131614Sdes		return x + efx*x;
213127094Sdes	    }
214176884Spiso	    z = x*x;
215131614Sdes	    r = pp0+z*(pp1+z*(pp2+z*(pp3+z*pp4)));
216176884Spiso	    s = one+z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5))));
217131614Sdes	    y = r/s;
218127094Sdes	    return x + x*y;
219127094Sdes	}
220176884Spiso	if(ix < 0x3ff40000) {		/* 0.84375 <= |x| < 1.25 */
221131614Sdes	    s = fabs(x)-one;
222127094Sdes	    P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))));
223127094Sdes	    Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))));
22426026Sbrian	    if(hx>=0) return erx + P/Q; else return -erx - P/Q;
22526026Sbrian	}
22626026Sbrian	if (ix >= 0x40180000) {		/* inf>|x|>=6 */
22726026Sbrian	    if(hx>=0) return one-tiny; else return tiny-one;
22826026Sbrian	}
22926026Sbrian	x = fabs(x);
23099207Sbrian 	s = one/(x*x);
23126026Sbrian	if(ix< 0x4006DB6E) {	/* |x| < 1/0.35 */
23265280Sru	    R=ra0+s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*(ra5+s*(ra6+s*ra7))))));
23365280Sru	    S=one+s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*(sa5+s*(sa6+s*(sa7+
23459726Sru		s*sa8)))))));
23526026Sbrian	} else {	/* |x| >= 1/0.35 */
23626026Sbrian	    R=rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(rb5+s*rb6)))));
23726026Sbrian	    S=one+s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*(sb5+s*(sb6+s*sb7))))));
23826026Sbrian	}
23926026Sbrian	z  = x;
24026026Sbrian	SET_LOW_WORD(z,0);
24126026Sbrian	r  =  __ieee754_exp(-z*z-0.5625)*__ieee754_exp((z-x)*(z+x)+R/S);
24226026Sbrian	if(hx>=0) return one-r/x; else return  r/x-one;
24326026Sbrian}
24426026Sbrian
24526026Sbriandouble
24626026Sbrianerfc(double x)
24726026Sbrian{
24826026Sbrian	int32_t hx,ix;
24959356Sru	double R,S,P,Q,s,y,z,r;
25026026Sbrian	GET_HIGH_WORD(hx,x);
25126026Sbrian	ix = hx&0x7fffffff;
25226026Sbrian	if(ix>=0x7ff00000) {			/* erfc(nan)=nan */
25326026Sbrian						/* erfc(+-inf)=0,2 */
25426026Sbrian	    return (double)(((u_int32_t)hx>>31)<<1)+one/x;
25526026Sbrian	}
25626026Sbrian
25726026Sbrian	if(ix < 0x3feb0000) {		/* |x|<0.84375 */
25859356Sru	    if(ix < 0x3c700000)  	/* |x|<2**-56 */
25926026Sbrian		return one-x;
26026026Sbrian	    z = x*x;
26126026Sbrian	    r = pp0+z*(pp1+z*(pp2+z*(pp3+z*pp4)));
26226026Sbrian	    s = one+z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5))));
26326026Sbrian	    y = r/s;
26426026Sbrian	    if(hx < 0x3fd00000) {  	/* x<1/4 */
26526026Sbrian		return one-(x+x*y);
26626026Sbrian	    } else {
26726026Sbrian		r = x*y;
26826026Sbrian		r += (x-half);
26926026Sbrian	        return half - r ;
27026026Sbrian	    }
271127094Sdes	}
272127094Sdes	if(ix < 0x3ff40000) {		/* 0.84375 <= |x| < 1.25 */
273127094Sdes	    s = fabs(x)-one;
27426026Sbrian	    P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))));
275133719Sphk	    Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))));
276127094Sdes	    if(hx>=0) {
277131566Sphk	        z  = one-erx; return z - P/Q;
27826026Sbrian	    } else {
279177098Spiso		z = erx+P/Q; return one+z;
280177098Spiso	    }
281177098Spiso	}
282177098Spiso	if (ix < 0x403c0000) {		/* |x|<28 */
283177098Spiso	    x = fabs(x);
28459726Sru 	    s = one/(x*x);
285127094Sdes	    if(ix< 0x4006DB6D) {	/* |x| < 1/.35 ~ 2.857143*/
286179920Smav		R=ra0+s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*(ra5+s*(ra6+s*ra7))))));
28726026Sbrian		S=one+s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*(sa5+s*(sa6+s*(sa7+
288127094Sdes		    s*sa8)))))));
289131566Sphk	    } else {			/* |x| >= 1/.35 ~ 2.857143 */
29026026Sbrian		if(hx<0&&ix>=0x40180000) return two-tiny;/* x < -6 */
29126026Sbrian		R=rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(rb5+s*rb6)))));
29226026Sbrian		S=one+s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*(sb5+s*(sb6+s*sb7))))));
293124621Sphk	    }
29426026Sbrian	    z  = x;
295165243Spiso	    SET_LOW_WORD(z,0);
296165243Spiso	    r  =  __ieee754_exp(-z*z-0.5625)*__ieee754_exp((z-x)*(z+x)+R/S);
29726026Sbrian	    if(hx>0) return r/x; else return two-r/x;
29865280Sru	} else {
29965280Sru	    if(hx>0) return tiny*tiny; else return two-tiny;
30026026Sbrian	}
301131614Sdes}
302127094Sdes