base/math/e_exp.c

Go to the documentation of this file.
00001 /* @(#)e_exp.c 5.1 93/09/24 */
00002 /*
00003  * ====================================================
00004  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
00005  *
00006  * Developed at SunPro, a Sun Microsystems, Inc. business.
00007  * Permission to use, copy, modify, and distribute this
00008  * software is freely granted, provided that this notice 
00009  * is preserved.
00010  * ====================================================
00011  */
00012 
00013 #if defined(LIBM_SCCS) && !defined(lint)
00014 static char rcsid[] = "$NetBSD: e_exp.c,v 1.8 1995/05/10 20:45:03 jtc Exp $";
00015 #endif
00016 
00017 /* __ieee754_exp(x)
00018  * Returns the exponential of x.
00019  *
00020  * Method
00021  *   1. Argument reduction:
00022  *      Reduce x to an r so that |r| <= 0.5*ln2 ~ 0.34658.
00023  *  Given x, find r and integer k such that
00024  *
00025  *               x = k*ln2 + r,  |r| <= 0.5*ln2.  
00026  *
00027  *      Here r will be represented as r = hi-lo for better 
00028  *  accuracy.
00029  *
00030  *   2. Approximation of exp(r) by a special rational function on
00031  *  the interval [0,0.34658]:
00032  *  Write
00033  *      R(r**2) = r*(exp(r)+1)/(exp(r)-1) = 2 + r*r/6 - r**4/360 + ...
00034  *      We use a special Reme algorithm on [0,0.34658] to generate 
00035  *  a polynomial of degree 5 to approximate R. The maximum error 
00036  *  of this polynomial approximation is bounded by 2**-59. In
00037  *  other words,
00038  *      R(z) ~ 2.0 + P1*z + P2*z**2 + P3*z**3 + P4*z**4 + P5*z**5
00039  *      (where z=r*r, and the values of P1 to P5 are listed below)
00040  *  and
00041  *      |                  5          |     -59
00042  *      | 2.0+P1*z+...+P5*z   -  R(z) | <= 2 
00043  *      |                             |
00044  *  The computation of exp(r) thus becomes
00045  *                             2*r
00046  *      exp(r) = 1 + -------
00047  *                    R - r
00048  *                                 r*R1(r)  
00049  *             = 1 + r + ----------- (for better accuracy)
00050  *                        2 - R1(r)
00051  *  where
00052  *                   2       4             10
00053  *      R1(r) = r - (P1*r  + P2*r  + ... + P5*r   ).
00054  *  
00055  *   3. Scale back to obtain exp(x):
00056  *  From step 1, we have
00057  *     exp(x) = 2^k * exp(r)
00058  *
00059  * Special cases:
00060  *  exp(INF) is INF, exp(NaN) is NaN;
00061  *  exp(-INF) is 0, and
00062  *  for finite argument, only exp(0)=1 is exact.
00063  *
00064  * Accuracy:
00065  *  according to an error analysis, the error is always less than
00066  *  1 ulp (unit in the last place).
00067  *
00068  * Misc. info.
00069  *  For IEEE double 
00070  *      if x >  7.09782712893383973096e+02 then exp(x) overflow
00071  *      if x < -7.45133219101941108420e+02 then exp(x) underflow
00072  *
00073  * Constants:
00074  * The hexadecimal values are the intended ones for the following 
00075  * constants. The decimal values may be used, provided that the 
00076  * compiler will convert from decimal to binary accurately enough
00077  * to produce the hexadecimal values shown.
00078  */
00079 
00080 #include "math.h"
00081 #include "mathP.h"
00082 
00083 #ifdef __STDC__
00084 static const double
00085 #else
00086 static double
00087 #endif
00088 one = 1.0,
00089 halF[2] = {0.5,-0.5,},
00090 huge    = 1.0e+300,
00091 twom1000= 9.33263618503218878990e-302,     /* 2**-1000=0x01700000,0*/
00092 o_threshold=  7.09782712893383973096e+02,  /* 0x40862E42, 0xFEFA39EF */
00093 u_threshold= -7.45133219101941108420e+02,  /* 0xc0874910, 0xD52D3051 */
00094 ln2HI[2]   ={ 6.93147180369123816490e-01,  /* 0x3fe62e42, 0xfee00000 */
00095          -6.93147180369123816490e-01,},/* 0xbfe62e42, 0xfee00000 */
00096 ln2LO[2]   ={ 1.90821492927058770002e-10,  /* 0x3dea39ef, 0x35793c76 */
00097          -1.90821492927058770002e-10,},/* 0xbdea39ef, 0x35793c76 */
00098 invln2 =  1.44269504088896338700e+00, /* 0x3ff71547, 0x652b82fe */
00099 P1   =  1.66666666666666019037e-01, /* 0x3FC55555, 0x5555553E */
00100 P2   = -2.77777777770155933842e-03, /* 0xBF66C16C, 0x16BEBD93 */
00101 P3   =  6.61375632143793436117e-05, /* 0x3F11566A, 0xAF25DE2C */
00102 P4   = -1.65339022054652515390e-06, /* 0xBEBBBD41, 0xC5D26BF1 */
00103 P5   =  4.13813679705723846039e-08; /* 0x3E663769, 0x72BEA4D0 */
00104 
00105 
00106 #ifdef __STDC__
00107     double __ieee754_exp(double x)  /* default IEEE double exp */
00108 #else
00109     double __ieee754_exp(x) /* default IEEE double exp */
00110     double x;
00111 #endif
00112 {
00113     double y,hi,lo,c,t;
00114     int32_t k,xsb;
00115     u_int32_t hx;
00116 
00117     /* XXX hi, lo, k are not properly initialized.  Bug?  --ds */
00118     hi = lo = k = 0;
00119 
00120     GET_HIGH_WORD(hx,x);
00121     xsb = (hx>>31)&1;       /* sign bit of x */
00122     hx &= 0x7fffffff;       /* high word of |x| */
00123 
00124     /* filter out non-finite argument */
00125     if(hx >= 0x40862E42) {          /* if |x|>=709.78... */
00126             if(hx>=0x7ff00000) {
00127             u_int32_t lx;
00128         GET_LOW_WORD(lx,x);
00129         if(((hx&0xfffff)|lx)!=0) 
00130              return x+x;        /* NaN */
00131         else return (xsb==0)? x:0.0;    /* exp(+-inf)={inf,0} */
00132         }
00133         if(x > o_threshold) return huge*huge; /* overflow */
00134         if(x < u_threshold) return twom1000*twom1000; /* underflow */
00135     }
00136 
00137     /* argument reduction */
00138     if(hx > 0x3fd62e42) {       /* if  |x| > 0.5 ln2 */ 
00139         if(hx < 0x3FF0A2B2) {   /* and |x| < 1.5 ln2 */
00140         hi = x-ln2HI[xsb]; lo=ln2LO[xsb]; k = 1-xsb-xsb;
00141         } else {
00142         k  = invln2*x+halF[xsb];
00143         t  = k;
00144         hi = x - t*ln2HI[0];    /* t*ln2HI is exact here */
00145         lo = t*ln2LO[0];
00146         }
00147         x  = hi - lo;
00148     } 
00149     else if(hx < 0x3e300000)  { /* when |x|<2**-28 */
00150         if(huge+x>one) return one+x;/* trigger inexact */
00151     }
00152     else k = 0;
00153 
00154     /* x is now in primary range */
00155     t  = x*x;
00156     c  = x - t*(P1+t*(P2+t*(P3+t*(P4+t*P5))));
00157     if(k==0)    return one-((x*c)/(c-2.0)-x); 
00158     else        y = one-((lo-(x*c)/(2.0-c))-hi);
00159     if(k >= -1021) {
00160         u_int32_t hy;
00161         GET_HIGH_WORD(hy,y);
00162         SET_HIGH_WORD(y,hy+(k<<20));    /* add k to y's exponent */
00163         return y;
00164     } else {
00165         u_int32_t hy;
00166         GET_HIGH_WORD(hy,y);
00167         SET_HIGH_WORD(y,hy+((k+1000)<<20)); /* add k to y's exponent */
00168         return y*twom1000;
00169     }
00170 }

Generated on Tue Feb 2 17:46:05 2010 for RTAI API by  doxygen 1.4.7