[388] | 1 | /* origin: FreeBSD /usr/src/lib/msun/src/e_exp.c */
|
---|
| 2 | /*
|
---|
| 3 | * ====================================================
|
---|
| 4 | * Copyright (C) 2004 by Sun Microsystems, Inc. All rights reserved.
|
---|
| 5 | *
|
---|
| 6 | * Permission to use, copy, modify, and distribute this
|
---|
| 7 | * software is freely granted, provided that this notice
|
---|
| 8 | * is preserved.
|
---|
| 9 | * ====================================================
|
---|
| 10 | */
|
---|
| 11 | /* exp(x)
|
---|
| 12 | * Returns the exponential of x.
|
---|
| 13 | *
|
---|
| 14 | * Method
|
---|
| 15 | * 1. Argument reduction:
|
---|
| 16 | * Reduce x to an r so that |r| <= 0.5*ln2 ~ 0.34658.
|
---|
| 17 | * Given x, find r and integer k such that
|
---|
| 18 | *
|
---|
| 19 | * x = k*ln2 + r, |r| <= 0.5*ln2.
|
---|
| 20 | *
|
---|
| 21 | * Here r will be represented as r = hi-lo for better
|
---|
| 22 | * accuracy.
|
---|
| 23 | *
|
---|
| 24 | * 2. Approximation of exp(r) by a special rational function on
|
---|
| 25 | * the interval [0,0.34658]:
|
---|
| 26 | * Write
|
---|
| 27 | * R(r**2) = r*(exp(r)+1)/(exp(r)-1) = 2 + r*r/6 - r**4/360 + ...
|
---|
| 28 | * We use a special Remez algorithm on [0,0.34658] to generate
|
---|
| 29 | * a polynomial of degree 5 to approximate R. The maximum error
|
---|
| 30 | * of this polynomial approximation is bounded by 2**-59. In
|
---|
| 31 | * other words,
|
---|
| 32 | * R(z) ~ 2.0 + P1*z + P2*z**2 + P3*z**3 + P4*z**4 + P5*z**5
|
---|
| 33 | * (where z=r*r, and the values of P1 to P5 are listed below)
|
---|
| 34 | * and
|
---|
| 35 | * | 5 | -59
|
---|
| 36 | * | 2.0+P1*z+...+P5*z - R(z) | <= 2
|
---|
| 37 | * | |
|
---|
| 38 | * The computation of exp(r) thus becomes
|
---|
| 39 | * 2*r
|
---|
| 40 | * exp(r) = 1 + ----------
|
---|
| 41 | * R(r) - r
|
---|
| 42 | * r*c(r)
|
---|
| 43 | * = 1 + r + ----------- (for better accuracy)
|
---|
| 44 | * 2 - c(r)
|
---|
| 45 | * where
|
---|
| 46 | * 2 4 10
|
---|
| 47 | * c(r) = r - (P1*r + P2*r + ... + P5*r ).
|
---|
| 48 | *
|
---|
| 49 | * 3. Scale back to obtain exp(x):
|
---|
| 50 | * From step 1, we have
|
---|
| 51 | * exp(x) = 2^k * exp(r)
|
---|
| 52 | *
|
---|
| 53 | * Special cases:
|
---|
| 54 | * exp(INF) is INF, exp(NaN) is NaN;
|
---|
| 55 | * exp(-INF) is 0, and
|
---|
| 56 | * for finite argument, only exp(0)=1 is exact.
|
---|
| 57 | *
|
---|
| 58 | * Accuracy:
|
---|
| 59 | * according to an error analysis, the error is always less than
|
---|
| 60 | * 1 ulp (unit in the last place).
|
---|
| 61 | *
|
---|
| 62 | * Misc. info.
|
---|
| 63 | * For IEEE double
|
---|
| 64 | * if x > 709.782712893383973096 then exp(x) overflows
|
---|
| 65 | * if x < -745.133219101941108420 then exp(x) underflows
|
---|
| 66 | */
|
---|
| 67 |
|
---|
| 68 | #include "libm.h"
|
---|
| 69 |
|
---|
| 70 | static const double
|
---|
| 71 | half[2] = {0.5,-0.5},
|
---|
| 72 | ln2hi = 6.93147180369123816490e-01, /* 0x3fe62e42, 0xfee00000 */
|
---|
| 73 | ln2lo = 1.90821492927058770002e-10, /* 0x3dea39ef, 0x35793c76 */
|
---|
| 74 | invln2 = 1.44269504088896338700e+00, /* 0x3ff71547, 0x652b82fe */
|
---|
| 75 | P1 = 1.66666666666666019037e-01, /* 0x3FC55555, 0x5555553E */
|
---|
| 76 | P2 = -2.77777777770155933842e-03, /* 0xBF66C16C, 0x16BEBD93 */
|
---|
| 77 | P3 = 6.61375632143793436117e-05, /* 0x3F11566A, 0xAF25DE2C */
|
---|
| 78 | P4 = -1.65339022054652515390e-06, /* 0xBEBBBD41, 0xC5D26BF1 */
|
---|
| 79 | P5 = 4.13813679705723846039e-08; /* 0x3E663769, 0x72BEA4D0 */
|
---|
| 80 |
|
---|
| 81 | double exp(double x)
|
---|
| 82 | {
|
---|
| 83 | double_t hi, lo, c, xx, y;
|
---|
| 84 | int k, sign;
|
---|
| 85 | uint32_t hx;
|
---|
| 86 |
|
---|
| 87 | GET_HIGH_WORD(hx, x);
|
---|
| 88 | sign = hx>>31;
|
---|
| 89 | hx &= 0x7fffffff; /* high word of |x| */
|
---|
| 90 |
|
---|
| 91 | /* special cases */
|
---|
| 92 | if (hx >= 0x4086232b) { /* if |x| >= 708.39... */
|
---|
| 93 | if (isnan(x))
|
---|
| 94 | return x;
|
---|
| 95 | if (x > 709.782712893383973096) {
|
---|
| 96 | /* overflow if x!=inf */
|
---|
| 97 | x *= 0x1p1023;
|
---|
| 98 | return x;
|
---|
| 99 | }
|
---|
| 100 | if (x < -708.39641853226410622) {
|
---|
| 101 | /* underflow if x!=-inf */
|
---|
| 102 | FORCE_EVAL((float)(-0x1p-149/x));
|
---|
| 103 | if (x < -745.13321910194110842)
|
---|
| 104 | return 0;
|
---|
| 105 | }
|
---|
| 106 | }
|
---|
| 107 |
|
---|
| 108 | /* argument reduction */
|
---|
| 109 | if (hx > 0x3fd62e42) { /* if |x| > 0.5 ln2 */
|
---|
| 110 | if (hx >= 0x3ff0a2b2) /* if |x| >= 1.5 ln2 */
|
---|
| 111 | k = (int)(invln2*x + half[sign]);
|
---|
| 112 | else
|
---|
| 113 | k = 1 - sign - sign;
|
---|
| 114 | hi = x - k*ln2hi; /* k*ln2hi is exact here */
|
---|
| 115 | lo = k*ln2lo;
|
---|
| 116 | x = hi - lo;
|
---|
| 117 | } else if (hx > 0x3e300000) { /* if |x| > 2**-28 */
|
---|
| 118 | k = 0;
|
---|
| 119 | hi = x;
|
---|
| 120 | lo = 0;
|
---|
| 121 | } else {
|
---|
| 122 | /* inexact if x!=0 */
|
---|
| 123 | FORCE_EVAL(0x1p1023 + x);
|
---|
| 124 | return 1 + x;
|
---|
| 125 | }
|
---|
| 126 |
|
---|
| 127 | /* x is now in primary range */
|
---|
| 128 | xx = x*x;
|
---|
| 129 | c = x - xx*(P1+xx*(P2+xx*(P3+xx*(P4+xx*P5))));
|
---|
| 130 | y = 1 + (x*c/(2-c) - lo + hi);
|
---|
| 131 | if (k == 0)
|
---|
| 132 | return y;
|
---|
| 133 | return scalbn(y, k);
|
---|
| 134 | }
|
---|