1 /* derived from /netlib/fdlibm */ 2 3 /* @(#)s_log1p.c 1.3 95/01/18 */ 4 /* 5 * ==================================================== 6 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. 7 * 8 * Developed at SunSoft, a Sun Microsystems, Inc. business. 9 * Permission to use, copy, modify, and distribute this 10 * software is freely granted, provided that this notice 11 * is preserved. 12 * ==================================================== 13 */ 14 15 /* double log1p(double x) 16 * 17 * Method : 18 * 1. Argument Reduction: find k and f such that 19 * 1+x = 2^k * (1+f), 20 * where sqrt(2)/2 < 1+f < sqrt(2) . 21 * 22 * Note. If k=0, then f=x is exact. However, if k!=0, then f 23 * may not be representable exactly. In that case, a correction 24 * term is need. Let u=1+x rounded. Let c = (1+x)-u, then 25 * log(1+x) - log(u) ~ c/u. Thus, we proceed to compute log(u), 26 * and add back the correction term c/u. 27 * (Note: when x > 2**53, one can simply return log(x)) 28 * 29 * 2. Approximation of log1p(f). 30 * Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s) 31 * = 2s + 2/3 s**3 + 2/5 s**5 + ....., 32 * = 2s + s*R 33 * We use a special Reme algorithm on [0,0.1716] to generate 34 * a polynomial of degree 14 to approximate R The maximum error 35 * of this polynomial approximation is bounded by 2**-58.45. In 36 * other words, 37 * 2 4 6 8 10 12 14 38 * R(z) ~ Lp1*s +Lp2*s +Lp3*s +Lp4*s +Lp5*s +Lp6*s +Lp7*s 39 * (the values of Lp1 to Lp7 are listed in the program) 40 * and 41 * | 2 14 | -58.45 42 * | Lp1*s +...+Lp7*s - R(z) | <= 2 43 * | | 44 * Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2. 45 * In order to guarantee error in log below 1ulp, we compute log 46 * by 47 * log1p(f) = f - (hfsq - s*(hfsq+R)). 48 * 49 * 3. Finally, log1p(x) = k*ln2 + log1p(f). 50 * = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo))) 51 * Here ln2 is split into two floating point number: 52 * ln2_hi + ln2_lo, 53 * where n*ln2_hi is always exact for |n| < 2000. 54 * 55 * Special cases: 56 * log1p(x) is NaN with signal if x < -1 (including -INF) ; 57 * log1p(+INF) is +INF; log1p(-1) is -INF with signal; 58 * log1p(NaN) is that NaN with no signal. 59 * 60 * Accuracy: 61 * according to an error analysis, the error is always less than 62 * 1 ulp (unit in the last place). 63 * 64 * Constants: 65 * The hexadecimal values are the intended ones for the following 66 * constants. The decimal values may be used, provided that the 67 * compiler will convert from decimal to binary accurately enough 68 * to produce the hexadecimal values shown. 69 * 70 * Note: Assuming log() return accurate answer, the following 71 * algorithm can be used to compute log1p(x) to within a few ULP: 72 * 73 * u = 1+x; 74 * if(u==1.0) return x ; else 75 * return log(u)*(x/(u-1.0)); 76 * 77 * See HP-15C Advanced Functions Handbook, p.193. 78 */ 79 80 #include "fdlibm.h" 81 82 static const double 83 ln2_hi = 6.93147180369123816490e-01, /* 3fe62e42 fee00000 */ 84 ln2_lo = 1.90821492927058770002e-10, /* 3dea39ef 35793c76 */ 85 two54 = 1.80143985094819840000e+16, /* 43500000 00000000 */ 86 Lp1 = 6.666666666666735130e-01, /* 3FE55555 55555593 */ 87 Lp2 = 3.999999999940941908e-01, /* 3FD99999 9997FA04 */ 88 Lp3 = 2.857142874366239149e-01, /* 3FD24924 94229359 */ 89 Lp4 = 2.222219843214978396e-01, /* 3FCC71C5 1D8E78AF */ 90 Lp5 = 1.818357216161805012e-01, /* 3FC74664 96CB03DE */ 91 Lp6 = 1.531383769920937332e-01, /* 3FC39A09 D078C69F */ 92 Lp7 = 1.479819860511658591e-01; /* 3FC2F112 DF3E5244 */ 93 94 static double zero = 0.0; 95 log1p(double x)96 double log1p(double x) 97 { 98 double hfsq,f,c,s,z,R,u; 99 int k,hx,hu,ax; 100 101 hx = __HI(x); /* high word of x */ 102 ax = hx&0x7fffffff; 103 104 k = 1; 105 if (hx < 0x3FDA827A) { /* x < 0.41422 */ 106 if(ax>=0x3ff00000) { /* x <= -1.0 */ 107 if(x==-1.0) return -two54/zero; /* log1p(-1)=+inf */ 108 else return (x-x)/(x-x); /* log1p(x<-1)=NaN */ 109 } 110 if(ax<0x3e200000) { /* |x| < 2**-29 */ 111 if(two54+x>zero /* raise inexact */ 112 &&ax<0x3c900000) /* |x| < 2**-54 */ 113 return x; 114 else 115 return x - x*x*0.5; 116 } 117 if(hx>0||hx<=((int)0xbfd2bec3)) { 118 k=0;f=x;hu=1;} /* -0.2929<x<0.41422 */ 119 } 120 if (hx >= 0x7ff00000) return x+x; 121 if(k!=0) { 122 if(hx<0x43400000) { 123 u = 1.0+x; 124 hu = __HI(u); /* high word of u */ 125 k = (hu>>20)-1023; 126 c = (k>0)? 1.0-(u-x):x-(u-1.0);/* correction term */ 127 c /= u; 128 } else { 129 u = x; 130 hu = __HI(u); /* high word of u */ 131 k = (hu>>20)-1023; 132 c = 0; 133 } 134 hu &= 0x000fffff; 135 if(hu<0x6a09e) { 136 __HI(u) = hu|0x3ff00000; /* normalize u */ 137 } else { 138 k += 1; 139 __HI(u) = hu|0x3fe00000; /* normalize u/2 */ 140 hu = (0x00100000-hu)>>2; 141 } 142 f = u-1.0; 143 } 144 hfsq=0.5*f*f; 145 if(hu==0) { /* |f| < 2**-20 */ 146 if(f==zero) if(k==0) return zero; 147 else {c += k*ln2_lo; return k*ln2_hi+c;} 148 R = hfsq*(1.0-0.66666666666666666*f); 149 if(k==0) return f-R; else 150 return k*ln2_hi-((R-(k*ln2_lo+c))-f); 151 } 152 s = f/(2.0+f); 153 z = s*s; 154 R = z*(Lp1+z*(Lp2+z*(Lp3+z*(Lp4+z*(Lp5+z*(Lp6+z*Lp7)))))); 155 if(k==0) return f-(hfsq-s*(hfsq+R)); else 156 return k*ln2_hi-((hfsq-(s*(hfsq+R)+(k*ln2_lo+c)))-f); 157 } 158