1 /* s_erff.c -- float version of s_erf.c.
2 * Conversion to float by Ian Lance Taylor, Cygnus Support, ian@cygnus.com.
3 */
4
5 /*
6 * ====================================================
7 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
8 *
9 * Developed at SunPro, a Sun Microsystems, Inc. business.
10 * Permission to use, copy, modify, and distribute this
11 * software is freely granted, provided that this notice
12 * is preserved.
13 * ====================================================
14 */
15
16 #include "math.h"
17 #include "math_private.h"
18
19 /* XXX Prevent compilers from erroneously constant folding: */
20 static const volatile float tiny = 1e-30;
21
22 static const float
23 half= 0.5,
24 one = 1,
25 two = 2,
26 erx = 8.42697144e-01, /* 0x3f57bb00 */
27 /*
28 * In the domain [0, 2**-14], only the first term in the power series
29 * expansion of erf(x) is used. The magnitude of the first neglected
30 * terms is less than 2**-42.
31 */
32 efx = 1.28379166e-01, /* 0x3e0375d4 */
33 efx8= 1.02703333e+00, /* 0x3f8375d4 */
34 /*
35 * Domain [0, 0.84375], range ~[-5.4419e-10, 5.5179e-10]:
36 * |(erf(x) - x)/x - pp(x)/qq(x)| < 2**-31
37 */
38 pp0 = 1.28379166e-01, /* 0x3e0375d4 */
39 pp1 = -3.36030394e-01, /* 0xbeac0c2d */
40 pp2 = -1.86261395e-03, /* 0xbaf422f4 */
41 qq1 = 3.12324315e-01, /* 0x3e9fe8f9 */
42 qq2 = 2.16070414e-02, /* 0x3cb10140 */
43 qq3 = -1.98859372e-03, /* 0xbb025311 */
44 /*
45 * Domain [0.84375, 1.25], range ~[-1.023e-9, 1.023e-9]:
46 * |(erf(x) - erx) - pa(x)/qa(x)| < 2**-31
47 */
48 pa0 = 3.65041046e-06, /* 0x3674f993 */
49 pa1 = 4.15109307e-01, /* 0x3ed48935 */
50 pa2 = -2.09395722e-01, /* 0xbe566bd5 */
51 pa3 = 8.67677554e-02, /* 0x3db1b34b */
52 qa1 = 4.95560974e-01, /* 0x3efdba2b */
53 qa2 = 3.71248513e-01, /* 0x3ebe1449 */
54 qa3 = 3.92478965e-02, /* 0x3d20c267 */
55 /*
56 * Domain [1.25,1/0.35], range ~[-4.821e-9, 4.927e-9]:
57 * |log(x*erfc(x)) + x**2 + 0.5625 - ra(x)/sa(x)| < 2**-28
58 */
59 ra0 = -9.88156721e-03, /* 0xbc21e64c */
60 ra1 = -5.43658376e-01, /* 0xbf0b2d32 */
61 ra2 = -1.66828310e+00, /* 0xbfd58a4d */
62 ra3 = -6.91554189e-01, /* 0xbf3109b2 */
63 sa1 = 4.48581553e+00, /* 0x408f8bcd */
64 sa2 = 4.10799170e+00, /* 0x408374ab */
65 sa3 = 5.53855181e-01, /* 0x3f0dc974 */
66 /*
67 * Domain [2.85715, 11], range ~[-1.484e-9, 1.505e-9]:
68 * |log(x*erfc(x)) + x**2 + 0.5625 - rb(x)/sb(x)| < 2**-30
69 */
70 rb0 = -9.86496918e-03, /* 0xbc21a0ae */
71 rb1 = -5.48049808e-01, /* 0xbf0c4cfe */
72 rb2 = -1.84115684e+00, /* 0xbfebab07 */
73 sb1 = 4.87132740e+00, /* 0x409be1ea */
74 sb2 = 3.04982710e+00, /* 0x4043305e */
75 sb3 = -7.61900663e-01; /* 0xbf430bec */
76
77 float
erff(float x)78 erff(float x)
79 {
80 int32_t hx,ix,i;
81 float R,S,P,Q,s,y,z,r;
82 GET_FLOAT_WORD(hx,x);
83 ix = hx&0x7fffffff;
84 if(ix>=0x7f800000) { /* erff(nan)=nan */
85 i = ((u_int32_t)hx>>31)<<1;
86 return (float)(1-i)+one/x; /* erff(+-inf)=+-1 */
87 }
88
89 if(ix < 0x3f580000) { /* |x|<0.84375 */
90 if(ix < 0x38800000) { /* |x|<2**-14 */
91 if (ix < 0x04000000) /* |x|<0x1p-119 */
92 return (8*x+efx8*x)/8; /* avoid spurious underflow */
93 return x + efx*x;
94 }
95 z = x*x;
96 r = pp0+z*(pp1+z*pp2);
97 s = one+z*(qq1+z*(qq2+z*qq3));
98 y = r/s;
99 return x + x*y;
100 }
101 if(ix < 0x3fa00000) { /* 0.84375 <= |x| < 1.25 */
102 s = fabsf(x)-one;
103 P = pa0+s*(pa1+s*(pa2+s*pa3));
104 Q = one+s*(qa1+s*(qa2+s*qa3));
105 if(hx>=0) return erx + P/Q; else return -erx - P/Q;
106 }
107 if (ix >= 0x40800000) { /* inf>|x|>=4 */
108 if(hx>=0) return one-tiny; else return tiny-one;
109 }
110 x = fabsf(x);
111 s = one/(x*x);
112 if(ix< 0x4036db8c) { /* |x| < 2.85715 ~ 1/0.35 */
113 R=ra0+s*(ra1+s*(ra2+s*ra3));
114 S=one+s*(sa1+s*(sa2+s*sa3));
115 } else { /* |x| >= 2.85715 ~ 1/0.35 */
116 R=rb0+s*(rb1+s*rb2);
117 S=one+s*(sb1+s*(sb2+s*sb3));
118 }
119 SET_FLOAT_WORD(z,hx&0xffffe000);
120 r = expf(-z*z-0.5625F)*expf((z-x)*(z+x)+R/S);
121 if(hx>=0) return one-r/x; else return r/x-one;
122 }
123
124 float
erfcf(float x)125 erfcf(float x)
126 {
127 int32_t hx,ix;
128 float R,S,P,Q,s,y,z,r;
129 GET_FLOAT_WORD(hx,x);
130 ix = hx&0x7fffffff;
131 if(ix>=0x7f800000) { /* erfcf(nan)=nan */
132 /* erfcf(+-inf)=0,2 */
133 return (float)(((u_int32_t)hx>>31)<<1)+one/x;
134 }
135
136 if(ix < 0x3f580000) { /* |x|<0.84375 */
137 if(ix < 0x33800000) /* |x|<2**-24 */
138 return one-x;
139 z = x*x;
140 r = pp0+z*(pp1+z*pp2);
141 s = one+z*(qq1+z*(qq2+z*qq3));
142 y = r/s;
143 if(hx < 0x3e800000) { /* x<1/4 */
144 return one-(x+x*y);
145 } else {
146 r = x*y;
147 r += (x-half);
148 return half - r ;
149 }
150 }
151 if(ix < 0x3fa00000) { /* 0.84375 <= |x| < 1.25 */
152 s = fabsf(x)-one;
153 P = pa0+s*(pa1+s*(pa2+s*pa3));
154 Q = one+s*(qa1+s*(qa2+s*qa3));
155 if(hx>=0) {
156 z = one-erx; return z - P/Q;
157 } else {
158 z = erx+P/Q; return one+z;
159 }
160 }
161 if (ix < 0x41300000) { /* |x|<11 */
162 x = fabsf(x);
163 s = one/(x*x);
164 if(ix< 0x4036db8c) { /* |x| < 2.85715 ~ 1/.35 */
165 R=ra0+s*(ra1+s*(ra2+s*ra3));
166 S=one+s*(sa1+s*(sa2+s*sa3));
167 } else { /* |x| >= 2.85715 ~ 1/.35 */
168 if(hx<0&&ix>=0x40a00000) return two-tiny;/* x < -5 */
169 R=rb0+s*(rb1+s*rb2);
170 S=one+s*(sb1+s*(sb2+s*sb3));
171 }
172 SET_FLOAT_WORD(z,hx&0xffffe000);
173 r = expf(-z*z-0.5625F)*expf((z-x)*(z+x)+R/S);
174 if(hx>0) return r/x; else return two-r/x;
175 } else {
176 if(hx>0) return tiny*tiny; else return two-tiny;
177 }
178 }
179