POK
erf.c
1 /*
2  * POK header
3  *
4  * The following file is a part of the POK project. Any modification should
5  * made according to the POK licence. You CANNOT use this file or a part of
6  * this file is this part of a file for your own project
7  *
8  * For more information on the POK licence, please see our LICENCE FILE
9  *
10  * Please follow the coding guidelines described in doc/CODING_GUIDELINES
11  *
12  * Copyright (c) 2007-2009 POK team
13  *
14  * Created by julien on Fri Jan 30 14:41:34 2009
15  */
16 
17 /* @(#)s_erf.c 5.1 93/09/24 */
18 /*
19  * ====================================================
20  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
21  *
22  * Developed at SunPro, a Sun Microsystems, Inc. business.
23  * Permission to use, copy, modify, and distribute this
24  * software is freely granted, provided that this notice
25  * is preserved.
26  * ====================================================
27  */
28 
29 /* double erf(double x)
30  * double erfc(double x)
31  * x
32  * 2 |\
33  * erf(x) = --------- | exp(-t*t)dt
34  * sqrt(pi) \|
35  * 0
36  *
37  * erfc(x) = 1-erf(x)
38  * Note that
39  * erf(-x) = -erf(x)
40  * erfc(-x) = 2 - erfc(x)
41  *
42  * Method:
43  * 1. For |x| in [0, 0.84375]
44  * erf(x) = x + x*R(x^2)
45  * erfc(x) = 1 - erf(x) if x in [-.84375,0.25]
46  * = 0.5 + ((0.5-x)-x*R) if x in [0.25,0.84375]
47  * where R = P/Q where P is an odd poly of degree 8 and
48  * Q is an odd poly of degree 10.
49  * -57.90
50  * | R - (erf(x)-x)/x | <= 2
51  *
52  *
53  * Remark. The formula is derived by noting
54  * erf(x) = (2/sqrt(pi))*(x - x^3/3 + x^5/10 - x^7/42 + ....)
55  * and that
56  * 2/sqrt(pi) = 1.128379167095512573896158903121545171688
57  * is close to one. The interval is chosen because the fix
58  * point of erf(x) is near 0.6174 (i.e., erf(x)=x when x is
59  * near 0.6174), and by some experiment, 0.84375 is chosen to
60  * guarantee the error is less than one ulp for erf.
61  *
62  * 2. For |x| in [0.84375,1.25], let s = |x| - 1, and
63  * c = 0.84506291151 rounded to single (24 bits)
64  * erf(x) = sign(x) * (c + P1(s)/Q1(s))
65  * erfc(x) = (1-c) - P1(s)/Q1(s) if x > 0
66  * 1+(c+P1(s)/Q1(s)) if x < 0
67  * |P1/Q1 - (erf(|x|)-c)| <= 2**-59.06
68  * Remark: here we use the taylor series expansion at x=1.
69  * erf(1+s) = erf(1) + s*Poly(s)
70  * = 0.845.. + P1(s)/Q1(s)
71  * That is, we use rational approximation to approximate
72  * erf(1+s) - (c = (single)0.84506291151)
73  * Note that |P1/Q1|< 0.078 for x in [0.84375,1.25]
74  * where
75  * P1(s) = degree 6 poly in s
76  * Q1(s) = degree 6 poly in s
77  *
78  * 3. For x in [1.25,1/0.35(~2.857143)],
79  * erfc(x) = (1/x)*exp(-x*x-0.5625+R1/S1)
80  * erf(x) = 1 - erfc(x)
81  * where
82  * R1(z) = degree 7 poly in z, (z=1/x^2)
83  * S1(z) = degree 8 poly in z
84  *
85  * 4. For x in [1/0.35,28]
86  * erfc(x) = (1/x)*exp(-x*x-0.5625+R2/S2) if x > 0
87  * = 2.0 - (1/x)*exp(-x*x-0.5625+R2/S2) if -6<x<0
88  * = 2.0 - tiny (if x <= -6)
89  * erf(x) = sign(x)*(1.0 - erfc(x)) if x < 6, else
90  * erf(x) = sign(x)*(1.0 - tiny)
91  * where
92  * R2(z) = degree 6 poly in z, (z=1/x^2)
93  * S2(z) = degree 7 poly in z
94  *
95  * Note1:
96  * To compute exp(-x*x-0.5625+R/S), let s be a single
97  * precision number and s := x; then
98  * -x*x = -s*s + (s-x)*(s+x)
99  * exp(-x*x-0.5626+R/S) =
100  * exp(-s*s-0.5625)*exp((s-x)*(s+x)+R/S);
101  * Note2:
102  * Here 4 and 5 make use of the asymptotic series
103  * exp(-x*x)
104  * erfc(x) ~ ---------- * ( 1 + Poly(1/x^2) )
105  * x*sqrt(pi)
106  * We use rational approximation to approximate
107  * g(s)=f(1/x^2) = log(erfc(x)*x) - x*x + 0.5625
108  * Here is the error bound for R1/S1 and R2/S2
109  * |R1/S1 - f(x)| < 2**(-62.57)
110  * |R2/S2 - f(x)| < 2**(-61.52)
111  *
112  * 5. For inf > x >= 28
113  * erf(x) = sign(x) *(1 - tiny) (raise inexact)
114  * erfc(x) = tiny*tiny (raise underflow) if x > 0
115  * = 2 - tiny if x<0
116  *
117  * 7. Special case:
118  * erf(0) = 0, erf(inf) = 1, erf(-inf) = -1,
119  * erfc(0) = 1, erfc(inf) = 0, erfc(-inf) = 2,
120  * erfc/erf(NaN) is NaN
121  */
122 
123 #ifdef POK_NEEDS_LIBMATH
124 #include <libm.h>
125 #include "math_private.h"
126 
127 static const double
128 tiny = 1e-300,
129 half= 5.00000000000000000000e-01, /* 0x3FE00000, 0x00000000 */
130 one = 1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */
131 two = 2.00000000000000000000e+00, /* 0x40000000, 0x00000000 */
132  /* c = (float)0.84506291151 */
133 erx = 8.45062911510467529297e-01, /* 0x3FEB0AC1, 0x60000000 */
134 /*
135  * Coefficients for approximation to erf on [0,0.84375]
136  */
137 efx = 1.28379167095512586316e-01, /* 0x3FC06EBA, 0x8214DB69 */
138 efx8= 1.02703333676410069053e+00, /* 0x3FF06EBA, 0x8214DB69 */
139 pp0 = 1.28379167095512558561e-01, /* 0x3FC06EBA, 0x8214DB68 */
140 pp1 = -3.25042107247001499370e-01, /* 0xBFD4CD7D, 0x691CB913 */
141 pp2 = -2.84817495755985104766e-02, /* 0xBF9D2A51, 0xDBD7194F */
142 pp3 = -5.77027029648944159157e-03, /* 0xBF77A291, 0x236668E4 */
143 pp4 = -2.37630166566501626084e-05, /* 0xBEF8EAD6, 0x120016AC */
144 qq1 = 3.97917223959155352819e-01, /* 0x3FD97779, 0xCDDADC09 */
145 qq2 = 6.50222499887672944485e-02, /* 0x3FB0A54C, 0x5536CEBA */
146 qq3 = 5.08130628187576562776e-03, /* 0x3F74D022, 0xC4D36B0F */
147 qq4 = 1.32494738004321644526e-04, /* 0x3F215DC9, 0x221C1A10 */
148 qq5 = -3.96022827877536812320e-06, /* 0xBED09C43, 0x42A26120 */
149 /*
150  * Coefficients for approximation to erf in [0.84375,1.25]
151  */
152 pa0 = -2.36211856075265944077e-03, /* 0xBF6359B8, 0xBEF77538 */
153 pa1 = 4.14856118683748331666e-01, /* 0x3FDA8D00, 0xAD92B34D */
154 pa2 = -3.72207876035701323847e-01, /* 0xBFD7D240, 0xFBB8C3F1 */
155 pa3 = 3.18346619901161753674e-01, /* 0x3FD45FCA, 0x805120E4 */
156 pa4 = -1.10894694282396677476e-01, /* 0xBFBC6398, 0x3D3E28EC */
157 pa5 = 3.54783043256182359371e-02, /* 0x3FA22A36, 0x599795EB */
158 pa6 = -2.16637559486879084300e-03, /* 0xBF61BF38, 0x0A96073F */
159 qa1 = 1.06420880400844228286e-01, /* 0x3FBB3E66, 0x18EEE323 */
160 qa2 = 5.40397917702171048937e-01, /* 0x3FE14AF0, 0x92EB6F33 */
161 qa3 = 7.18286544141962662868e-02, /* 0x3FB2635C, 0xD99FE9A7 */
162 qa4 = 1.26171219808761642112e-01, /* 0x3FC02660, 0xE763351F */
163 qa5 = 1.36370839120290507362e-02, /* 0x3F8BEDC2, 0x6B51DD1C */
164 qa6 = 1.19844998467991074170e-02, /* 0x3F888B54, 0x5735151D */
165 /*
166  * Coefficients for approximation to erfc in [1.25,1/0.35]
167  */
168 ra0 = -9.86494403484714822705e-03, /* 0xBF843412, 0x600D6435 */
169 ra1 = -6.93858572707181764372e-01, /* 0xBFE63416, 0xE4BA7360 */
170 ra2 = -1.05586262253232909814e+01, /* 0xC0251E04, 0x41B0E726 */
171 ra3 = -6.23753324503260060396e+01, /* 0xC04F300A, 0xE4CBA38D */
172 ra4 = -1.62396669462573470355e+02, /* 0xC0644CB1, 0x84282266 */
173 ra5 = -1.84605092906711035994e+02, /* 0xC067135C, 0xEBCCABB2 */
174 ra6 = -8.12874355063065934246e+01, /* 0xC0545265, 0x57E4D2F2 */
175 ra7 = -9.81432934416914548592e+00, /* 0xC023A0EF, 0xC69AC25C */
176 sa1 = 1.96512716674392571292e+01, /* 0x4033A6B9, 0xBD707687 */
177 sa2 = 1.37657754143519042600e+02, /* 0x4061350C, 0x526AE721 */
178 sa3 = 4.34565877475229228821e+02, /* 0x407B290D, 0xD58A1A71 */
179 sa4 = 6.45387271733267880336e+02, /* 0x40842B19, 0x21EC2868 */
180 sa5 = 4.29008140027567833386e+02, /* 0x407AD021, 0x57700314 */
181 sa6 = 1.08635005541779435134e+02, /* 0x405B28A3, 0xEE48AE2C */
182 sa7 = 6.57024977031928170135e+00, /* 0x401A47EF, 0x8E484A93 */
183 sa8 = -6.04244152148580987438e-02, /* 0xBFAEEFF2, 0xEE749A62 */
184 /*
185  * Coefficients for approximation to erfc in [1/.35,28]
186  */
187 rb0 = -9.86494292470009928597e-03, /* 0xBF843412, 0x39E86F4A */
188 rb1 = -7.99283237680523006574e-01, /* 0xBFE993BA, 0x70C285DE */
189 rb2 = -1.77579549177547519889e+01, /* 0xC031C209, 0x555F995A */
190 rb3 = -1.60636384855821916062e+02, /* 0xC064145D, 0x43C5ED98 */
191 rb4 = -6.37566443368389627722e+02, /* 0xC083EC88, 0x1375F228 */
192 rb5 = -1.02509513161107724954e+03, /* 0xC0900461, 0x6A2E5992 */
193 rb6 = -4.83519191608651397019e+02, /* 0xC07E384E, 0x9BDC383F */
194 sb1 = 3.03380607434824582924e+01, /* 0x403E568B, 0x261D5190 */
195 sb2 = 3.25792512996573918826e+02, /* 0x40745CAE, 0x221B9F0A */
196 sb3 = 1.53672958608443695994e+03, /* 0x409802EB, 0x189D5118 */
197 sb4 = 3.19985821950859553908e+03, /* 0x40A8FFB7, 0x688C246A */
198 sb5 = 2.55305040643316442583e+03, /* 0x40A3F219, 0xCEDF3BE6 */
199 sb6 = 4.74528541206955367215e+02, /* 0x407DA874, 0xE79FE763 */
200 sb7 = -2.24409524465858183362e+01; /* 0xC03670E2, 0x42712D62 */
201 
202 double
203 erf(double x)
204 {
205  int32_t hx,ix,i;
206  double R,S,P,Q,s,y,z,r;
207  GET_HIGH_WORD(hx,x);
208  ix = hx&0x7fffffff;
209  if(ix>=0x7ff00000) { /* erf(nan)=nan */
210  i = ((uint32_t)hx>>31)<<1;
211  return (double)(1-i)+one/x; /* erf(+-inf)=+-1 */
212  }
213 
214  if(ix < 0x3feb0000) { /* |x|<0.84375 */
215  if(ix < 0x3e300000) { /* |x|<2**-28 */
216  if (ix < 0x00800000)
217  return 0.125*(8.0*x+efx8*x); /*avoid underflow */
218  return x + efx*x;
219  }
220  z = x*x;
221  r = pp0+z*(pp1+z*(pp2+z*(pp3+z*pp4)));
222  s = one+z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5))));
223  y = r/s;
224  return x + x*y;
225  }
226  if(ix < 0x3ff40000) { /* 0.84375 <= |x| < 1.25 */
227  s = fabs(x)-one;
228  P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))));
229  Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))));
230  if(hx>=0) return erx + P/Q; else return -erx - P/Q;
231  }
232  if (ix >= 0x40180000) { /* inf>|x|>=6 */
233  if(hx>=0) return one-tiny; else return tiny-one;
234  }
235  x = fabs(x);
236  s = one/(x*x);
237  if(ix< 0x4006DB6E) { /* |x| < 1/0.35 */
238  R=ra0+s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*(
239  ra5+s*(ra6+s*ra7))))));
240  S=one+s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*(
241  sa5+s*(sa6+s*(sa7+s*sa8)))))));
242  } else { /* |x| >= 1/0.35 */
243  R=rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(
244  rb5+s*rb6)))));
245  S=one+s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*(
246  sb5+s*(sb6+s*sb7))))));
247  }
248  z = x;
249  SET_LOW_WORD(z,0);
250  r = __ieee754_exp(-z*z-0.5625)*__ieee754_exp((z-x)*(z+x)+R/S);
251  if(hx>=0) return one-r/x; else return r/x-one;
252 }
253 
254 double
255 erfc(double x)
256 {
257  int32_t hx,ix;
258  double R,S,P,Q,s,y,z,r;
259  GET_HIGH_WORD(hx,x);
260  ix = hx&0x7fffffff;
261  if(ix>=0x7ff00000) { /* erfc(nan)=nan */
262  /* erfc(+-inf)=0,2 */
263  return (double)(((uint32_t)hx>>31)<<1)+one/x;
264  }
265 
266  if(ix < 0x3feb0000) { /* |x|<0.84375 */
267  if(ix < 0x3c700000) /* |x|<2**-56 */
268  return one-x;
269  z = x*x;
270  r = pp0+z*(pp1+z*(pp2+z*(pp3+z*pp4)));
271  s = one+z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5))));
272  y = r/s;
273  if(hx < 0x3fd00000) { /* x<1/4 */
274  return one-(x+x*y);
275  } else {
276  r = x*y;
277  r += (x-half);
278  return half - r ;
279  }
280  }
281  if(ix < 0x3ff40000) { /* 0.84375 <= |x| < 1.25 */
282  s = fabs(x)-one;
283  P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))));
284  Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))));
285  if(hx>=0) {
286  z = one-erx; return z - P/Q;
287  } else {
288  z = erx+P/Q; return one+z;
289  }
290  }
291  if (ix < 0x403c0000) { /* |x|<28 */
292  x = fabs(x);
293  s = one/(x*x);
294  if(ix< 0x4006DB6D) { /* |x| < 1/.35 ~ 2.857143*/
295  R=ra0+s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*(
296  ra5+s*(ra6+s*ra7))))));
297  S=one+s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*(
298  sa5+s*(sa6+s*(sa7+s*sa8)))))));
299  } else { /* |x| >= 1/.35 ~ 2.857143 */
300  if(hx<0&&ix>=0x40180000) return two-tiny;/* x < -6 */
301  R=rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(
302  rb5+s*rb6)))));
303  S=one+s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*(
304  sb5+s*(sb6+s*sb7))))));
305  }
306  z = x;
307  SET_LOW_WORD(z,0);
308  r = __ieee754_exp(-z*z-0.5625)*
309  __ieee754_exp((z-x)*(z+x)+R/S);
310  if(hx>0) return r/x; else return two-r/x;
311  } else {
312  if(hx>0) return tiny*tiny; else return two-tiny;
313  }
314 }
315 #endif