1 2 /* @(#)e_exp.c 1.6 04/04/22 */ 3 /* 4 * ==================================================== 5 * Copyright (C) 2004 by Sun Microsystems, Inc. All rights reserved. 6 * 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 /* __ieee754_exp(x) 14 * Returns the exponential of x. 15 * 16 * Method 17 * 1. Argument reduction: 18 * Reduce x to an r so that |r| <= 0.5*ln2 ~ 0.34658. 19 * Given x, find r and integer k such that 20 * 21 * x = k*ln2 + r, |r| <= 0.5*ln2. 22 * 23 * Here r will be represented as r = hi-lo for better 24 * accuracy. 25 * 26 * 2. Approximation of ieee_exp(r) by a special rational function on 27 * the interval [0,0.34658]: 28 * Write 29 * R(r**2) = r*(ieee_exp(r)+1)/(ieee_exp(r)-1) = 2 + r*r/6 - r**4/360 + ... 30 * We use a special Remes algorithm on [0,0.34658] to generate 31 * a polynomial of degree 5 to approximate R. The maximum error 32 * of this polynomial approximation is bounded by 2**-59. In 33 * other words, 34 * R(z) ~ 2.0 + P1*z + P2*z**2 + P3*z**3 + P4*z**4 + P5*z**5 35 * (where z=r*r, and the values of P1 to P5 are listed below) 36 * and 37 * | 5 | -59 38 * | 2.0+P1*z+...+P5*z - R(z) | <= 2 39 * | | 40 * The computation of ieee_exp(r) thus becomes 41 * 2*r 42 * exp(r) = 1 + ------- 43 * R - r 44 * r*R1(r) 45 * = 1 + r + ----------- (for better accuracy) 46 * 2 - R1(r) 47 * where 48 * 2 4 10 49 * R1(r) = r - (P1*r + P2*r + ... + P5*r ). 50 * 51 * 3. Scale back to obtain ieee_exp(x): 52 * From step 1, we have 53 * ieee_exp(x) = 2^k * ieee_exp(r) 54 * 55 * Special cases: 56 * exp(INF) is INF, ieee_exp(NaN) is NaN; 57 * exp(-INF) is 0, and 58 * for finite argument, only ieee_exp(0)=1 is exact. 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 * Misc. info. 65 * For IEEE double 66 * if x > 7.09782712893383973096e+02 then ieee_exp(x) overflow 67 * if x < -7.45133219101941108420e+02 then ieee_exp(x) underflow 68 * 69 * Constants: 70 * The hexadecimal values are the intended ones for the following 71 * constants. The decimal values may be used, provided that the 72 * compiler will convert from decimal to binary accurately enough 73 * to produce the hexadecimal values shown. 74 */ 75 76 #include "fdlibm.h" 77 78 #ifdef __STDC__ 79 static const double 80 #else 81 static double 82 #endif 83 one = 1.0, 84 halF[2] = {0.5,-0.5,}, 85 huge = 1.0e+300, 86 twom1000= 9.33263618503218878990e-302, /* 2**-1000=0x01700000,0*/ 87 o_threshold= 7.09782712893383973096e+02, /* 0x40862E42, 0xFEFA39EF */ 88 u_threshold= -7.45133219101941108420e+02, /* 0xc0874910, 0xD52D3051 */ 89 ln2HI[2] ={ 6.93147180369123816490e-01, /* 0x3fe62e42, 0xfee00000 */ 90 -6.93147180369123816490e-01,},/* 0xbfe62e42, 0xfee00000 */ 91 ln2LO[2] ={ 1.90821492927058770002e-10, /* 0x3dea39ef, 0x35793c76 */ 92 -1.90821492927058770002e-10,},/* 0xbdea39ef, 0x35793c76 */ 93 invln2 = 1.44269504088896338700e+00, /* 0x3ff71547, 0x652b82fe */ 94 P1 = 1.66666666666666019037e-01, /* 0x3FC55555, 0x5555553E */ 95 P2 = -2.77777777770155933842e-03, /* 0xBF66C16C, 0x16BEBD93 */ 96 P3 = 6.61375632143793436117e-05, /* 0x3F11566A, 0xAF25DE2C */ 97 P4 = -1.65339022054652515390e-06, /* 0xBEBBBD41, 0xC5D26BF1 */ 98 P5 = 4.13813679705723846039e-08; /* 0x3E663769, 0x72BEA4D0 */ 99 100 101 #ifdef __STDC__ __ieee754_exp(double x)102 double __ieee754_exp(double x) /* default IEEE double exp */ 103 #else 104 double __ieee754_exp(x) /* default IEEE double exp */ 105 double x; 106 #endif 107 { 108 double y,hi,lo,c,t; 109 int k,xsb; 110 unsigned hx; 111 112 hx = __HI(x); /* high word of x */ 113 xsb = (hx>>31)&1; /* sign bit of x */ 114 hx &= 0x7fffffff; /* high word of |x| */ 115 116 /* filter out non-finite argument */ 117 if(hx >= 0x40862E42) { /* if |x|>=709.78... */ 118 if(hx>=0x7ff00000) { 119 if(((hx&0xfffff)|__LO(x))!=0) 120 return x+x; /* NaN */ 121 else return (xsb==0)? x:0.0; /* ieee_exp(+-inf)={inf,0} */ 122 } 123 if(x > o_threshold) return huge*huge; /* overflow */ 124 if(x < u_threshold) return twom1000*twom1000; /* underflow */ 125 } 126 127 /* argument reduction */ 128 if(hx > 0x3fd62e42) { /* if |x| > 0.5 ln2 */ 129 if(hx < 0x3FF0A2B2) { /* and |x| < 1.5 ln2 */ 130 hi = x-ln2HI[xsb]; lo=ln2LO[xsb]; k = 1-xsb-xsb; 131 } else { 132 k = (int)(invln2*x+halF[xsb]); 133 t = k; 134 hi = x - t*ln2HI[0]; /* t*ln2HI is exact here */ 135 lo = t*ln2LO[0]; 136 } 137 x = hi - lo; 138 } 139 else if(hx < 0x3e300000) { /* when |x|<2**-28 */ 140 if(huge+x>one) return one+x;/* trigger inexact */ 141 } 142 else k = 0; 143 144 /* x is now in primary range */ 145 t = x*x; 146 c = x - t*(P1+t*(P2+t*(P3+t*(P4+t*P5)))); 147 if(k==0) return one-((x*c)/(c-2.0)-x); 148 else y = one-((lo-(x*c)/(2.0-c))-hi); 149 if(k >= -1021) { 150 __HI(y) += (k<<20); /* add k to y's exponent */ 151 return y; 152 } else { 153 __HI(y) += ((k+1000)<<20);/* add k to y's exponent */ 154 return y*twom1000; 155 } 156 } 157