vx32

Local 9vx git repository for patches.
git clone git://r-36.net/vx32
Log | Files | Refs

s_log1p.c (5264B)


      1 /* @(#)s_log1p.c 5.1 93/09/24 */
      2 /*
      3  * ====================================================
      4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
      5  *
      6  * Developed at SunPro, a Sun Microsystems, Inc. business.
      7  * Permission to use, copy, modify, and distribute this
      8  * software is freely granted, provided that this notice
      9  * is preserved.
     10  * ====================================================
     11  */
     12 
     13 #ifndef lint
     14 static char rcsid[] = "$FreeBSD: src/lib/msun/src/s_log1p.c,v 1.7 2002/05/28 18:15:04 alfred Exp $";
     15 #endif
     16 
     17 /* double log1p(double x)
     18  *
     19  * Method :
     20  *   1. Argument Reduction: find k and f such that
     21  *			1+x = 2^k * (1+f),
     22  *	   where  sqrt(2)/2 < 1+f < sqrt(2) .
     23  *
     24  *      Note. If k=0, then f=x is exact. However, if k!=0, then f
     25  *	may not be representable exactly. In that case, a correction
     26  *	term is need. Let u=1+x rounded. Let c = (1+x)-u, then
     27  *	log(1+x) - log(u) ~ c/u. Thus, we proceed to compute log(u),
     28  *	and add back the correction term c/u.
     29  *	(Note: when x > 2**53, one can simply return log(x))
     30  *
     31  *   2. Approximation of log1p(f).
     32  *	Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
     33  *		 = 2s + 2/3 s**3 + 2/5 s**5 + .....,
     34  *	     	 = 2s + s*R
     35  *      We use a special Reme algorithm on [0,0.1716] to generate
     36  * 	a polynomial of degree 14 to approximate R The maximum error
     37  *	of this polynomial approximation is bounded by 2**-58.45. In
     38  *	other words,
     39  *		        2      4      6      8      10      12      14
     40  *	    R(z) ~ Lp1*s +Lp2*s +Lp3*s +Lp4*s +Lp5*s  +Lp6*s  +Lp7*s
     41  *  	(the values of Lp1 to Lp7 are listed in the program)
     42  *	and
     43  *	    |      2          14          |     -58.45
     44  *	    | Lp1*s +...+Lp7*s    -  R(z) | <= 2
     45  *	    |                             |
     46  *	Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2.
     47  *	In order to guarantee error in log below 1ulp, we compute log
     48  *	by
     49  *		log1p(f) = f - (hfsq - s*(hfsq+R)).
     50  *
     51  *	3. Finally, log1p(x) = k*ln2 + log1p(f).
     52  *		 	     = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo)))
     53  *	   Here ln2 is split into two floating point number:
     54  *			ln2_hi + ln2_lo,
     55  *	   where n*ln2_hi is always exact for |n| < 2000.
     56  *
     57  * Special cases:
     58  *	log1p(x) is NaN with signal if x < -1 (including -INF) ;
     59  *	log1p(+INF) is +INF; log1p(-1) is -INF with signal;
     60  *	log1p(NaN) is that NaN with no signal.
     61  *
     62  * Accuracy:
     63  *	according to an error analysis, the error is always less than
     64  *	1 ulp (unit in the last place).
     65  *
     66  * Constants:
     67  * The hexadecimal values are the intended ones for the following
     68  * constants. The decimal values may be used, provided that the
     69  * compiler will convert from decimal to binary accurately enough
     70  * to produce the hexadecimal values shown.
     71  *
     72  * Note: Assuming log() return accurate answer, the following
     73  * 	 algorithm can be used to compute log1p(x) to within a few ULP:
     74  *
     75  *		u = 1+x;
     76  *		if(u==1.0) return x ; else
     77  *			   return log(u)*(x/(u-1.0));
     78  *
     79  *	 See HP-15C Advanced Functions Handbook, p.193.
     80  */
     81 
     82 #include "math.h"
     83 #include "math_private.h"
     84 
     85 static const double
     86 ln2_hi  =  6.93147180369123816490e-01,	/* 3fe62e42 fee00000 */
     87 ln2_lo  =  1.90821492927058770002e-10,	/* 3dea39ef 35793c76 */
     88 two54   =  1.80143985094819840000e+16,  /* 43500000 00000000 */
     89 Lp1 = 6.666666666666735130e-01,  /* 3FE55555 55555593 */
     90 Lp2 = 3.999999999940941908e-01,  /* 3FD99999 9997FA04 */
     91 Lp3 = 2.857142874366239149e-01,  /* 3FD24924 94229359 */
     92 Lp4 = 2.222219843214978396e-01,  /* 3FCC71C5 1D8E78AF */
     93 Lp5 = 1.818357216161805012e-01,  /* 3FC74664 96CB03DE */
     94 Lp6 = 1.531383769920937332e-01,  /* 3FC39A09 D078C69F */
     95 Lp7 = 1.479819860511658591e-01;  /* 3FC2F112 DF3E5244 */
     96 
     97 static const double zero = 0.0;
     98 
     99 double
    100 log1p(double x)
    101 {
    102 	double hfsq,f,c,s,z,R,u;
    103 	int32_t k,hx,hu,ax;
    104 
    105 	GET_HIGH_WORD(hx,x);
    106 	ax = hx&0x7fffffff;
    107 
    108 	k = 1;
    109 	if (hx < 0x3FDA827A) {			/* x < 0.41422  */
    110 	    if(ax>=0x3ff00000) {		/* x <= -1.0 */
    111 		if(x==-1.0) return -two54/zero; /* log1p(-1)=+inf */
    112 		else return (x-x)/(x-x);	/* log1p(x<-1)=NaN */
    113 	    }
    114 	    if(ax<0x3e200000) {			/* |x| < 2**-29 */
    115 		if(two54+x>zero			/* raise inexact */
    116 	            &&ax<0x3c900000) 		/* |x| < 2**-54 */
    117 		    return x;
    118 		else
    119 		    return x - x*x*0.5;
    120 	    }
    121 	    if(hx>0||hx<=((int32_t)0xbfd2bec3)) {
    122 		k=0;f=x;hu=1;}	/* -0.2929<x<0.41422 */
    123 	}
    124 	if (hx >= 0x7ff00000) return x+x;
    125 	if(k!=0) {
    126 	    if(hx<0x43400000) {
    127 		u  = 1.0+x;
    128 		GET_HIGH_WORD(hu,u);
    129 	        k  = (hu>>20)-1023;
    130 	        c  = (k>0)? 1.0-(u-x):x-(u-1.0);/* correction term */
    131 		c /= u;
    132 	    } else {
    133 		u  = x;
    134 		GET_HIGH_WORD(hu,u);
    135 	        k  = (hu>>20)-1023;
    136 		c  = 0;
    137 	    }
    138 	    hu &= 0x000fffff;
    139 	    if(hu<0x6a09e) {
    140 	        SET_HIGH_WORD(u,hu|0x3ff00000);	/* normalize u */
    141 	    } else {
    142 	        k += 1;
    143 		SET_HIGH_WORD(u,hu|0x3fe00000);	/* normalize u/2 */
    144 	        hu = (0x00100000-hu)>>2;
    145 	    }
    146 	    f = u-1.0;
    147 	}
    148 	hfsq=0.5*f*f;
    149 	if(hu==0) {	/* |f| < 2**-20 */
    150 	    if(f==zero) if(k==0) return zero;
    151 			else {c += k*ln2_lo; return k*ln2_hi+c;}
    152 	    R = hfsq*(1.0-0.66666666666666666*f);
    153 	    if(k==0) return f-R; else
    154 	    	     return k*ln2_hi-((R-(k*ln2_lo+c))-f);
    155 	}
    156  	s = f/(2.0+f);
    157 	z = s*s;
    158 	R = z*(Lp1+z*(Lp2+z*(Lp3+z*(Lp4+z*(Lp5+z*(Lp6+z*Lp7))))));
    159 	if(k==0) return f-(hfsq-s*(hfsq+R)); else
    160 		 return k*ln2_hi-((hfsq-(s*(hfsq+R)+(k*ln2_lo+c)))-f);
    161 }