|
| 1 | +/* origin: FreeBSD /usr/src/lib/msun/src/e_log.c */ |
| 2 | +/* |
| 3 | + * ==================================================== |
| 4 | + * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
| 5 | + * |
| 6 | + * Developed at SunSoft, 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 | +/* log(x) |
| 13 | + * Return the logarithm of x |
| 14 | + * |
| 15 | + * Method : |
| 16 | + * 1. Argument Reduction: find k and f such that |
| 17 | + * x = 2^k * (1+f), |
| 18 | + * where sqrt(2)/2 < 1+f < sqrt(2) . |
| 19 | + * |
| 20 | + * 2. Approximation of log(1+f). |
| 21 | + * Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s) |
| 22 | + * = 2s + 2/3 s**3 + 2/5 s**5 + ....., |
| 23 | + * = 2s + s*R |
| 24 | + * We use a special Remez algorithm on [0,0.1716] to generate |
| 25 | + * a polynomial of degree 14 to approximate R The maximum error |
| 26 | + * of this polynomial approximation is bounded by 2**-58.45. In |
| 27 | + * other words, |
| 28 | + * 2 4 6 8 10 12 14 |
| 29 | + * R(z) ~ Lg1*s +Lg2*s +Lg3*s +Lg4*s +Lg5*s +Lg6*s +Lg7*s |
| 30 | + * (the values of Lg1 to Lg7 are listed in the program) |
| 31 | + * and |
| 32 | + * | 2 14 | -58.45 |
| 33 | + * | Lg1*s +...+Lg7*s - R(z) | <= 2 |
| 34 | + * | | |
| 35 | + * Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2. |
| 36 | + * In order to guarantee error in log below 1ulp, we compute log |
| 37 | + * by |
| 38 | + * log(1+f) = f - s*(f - R) (if f is not too large) |
| 39 | + * log(1+f) = f - (hfsq - s*(hfsq+R)). (better accuracy) |
| 40 | + * |
| 41 | + * 3. Finally, log(x) = k*ln2 + log(1+f). |
| 42 | + * = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo))) |
| 43 | + * Here ln2 is split into two floating point number: |
| 44 | + * ln2_hi + ln2_lo, |
| 45 | + * where n*ln2_hi is always exact for |n| < 2000. |
| 46 | + * |
| 47 | + * Special cases: |
| 48 | + * log(x) is NaN with signal if x < 0 (including -INF) ; |
| 49 | + * log(+INF) is +INF; log(0) is -INF with signal; |
| 50 | + * log(NaN) is that NaN with no signal. |
| 51 | + * |
| 52 | + * Accuracy: |
| 53 | + * according to an error analysis, the error is always less than |
| 54 | + * 1 ulp (unit in the last place). |
| 55 | + * |
| 56 | + * Constants: |
| 57 | + * The hexadecimal values are the intended ones for the following |
| 58 | + * constants. The decimal values may be used, provided that the |
| 59 | + * compiler will convert from decimal to binary accurately enough |
| 60 | + * to produce the hexadecimal values shown. |
| 61 | + */ |
| 62 | + |
| 63 | +const LN2_HI: f64 = 6.93147180369123816490e-01; /* 3fe62e42 fee00000 */ |
| 64 | +const LN2_LO: f64 = 1.90821492927058770002e-10; /* 3dea39ef 35793c76 */ |
| 65 | +const LG1: f64 = 6.666666666666735130e-01; /* 3FE55555 55555593 */ |
| 66 | +const LG2: f64 = 3.999999999940941908e-01; /* 3FD99999 9997FA04 */ |
| 67 | +const LG3: f64 = 2.857142874366239149e-01; /* 3FD24924 94229359 */ |
| 68 | +const LG4: f64 = 2.222219843214978396e-01; /* 3FCC71C5 1D8E78AF */ |
| 69 | +const LG5: f64 = 1.818357216161805012e-01; /* 3FC74664 96CB03DE */ |
| 70 | +const LG6: f64 = 1.531383769920937332e-01; /* 3FC39A09 D078C69F */ |
| 71 | +const LG7: f64 = 1.479819860511658591e-01; /* 3FC2F112 DF3E5244 */ |
| 72 | + |
| 73 | +#[inline] |
| 74 | +pub fn log(mut x: f64) -> f64 { |
| 75 | + let x1p54 = f64::from_bits(0x4350000000000000); // 0x1p54 === 2 ^ 54 |
| 76 | + |
| 77 | + let mut ui = x.to_bits(); |
| 78 | + let mut hx: u32 = (ui >> 32) as u32; |
| 79 | + let mut k: i32 = 0; |
| 80 | + |
| 81 | + if (hx < 0x00100000) || ((hx >> 31) != 0) { |
| 82 | + /* x < 2**-126 */ |
| 83 | + if ui << 1 == 0 { |
| 84 | + return -1. / (x * x); /* log(+-0)=-inf */ |
| 85 | + } |
| 86 | + if hx >> 31 != 0 { |
| 87 | + return (x - x) / 0.0; /* log(-#) = NaN */ |
| 88 | + } |
| 89 | + /* subnormal number, scale x up */ |
| 90 | + k -= 54; |
| 91 | + x *= x1p54; |
| 92 | + ui = x.to_bits(); |
| 93 | + hx = (ui >> 32) as u32; |
| 94 | + } else if hx >= 0x7ff00000 { |
| 95 | + return x; |
| 96 | + } else if hx == 0x3ff00000 && ui << 32 == 0 { |
| 97 | + return 0.; |
| 98 | + } |
| 99 | + |
| 100 | + /* reduce x into [sqrt(2)/2, sqrt(2)] */ |
| 101 | + hx += 0x3ff00000 - 0x3fe6a09e; |
| 102 | + k += ((hx >> 20) as i32) - 0x3ff; |
| 103 | + hx = (hx & 0x000fffff) + 0x3fe6a09e; |
| 104 | + ui = ((hx as u64) << 32) | (ui & 0xffffffff); |
| 105 | + x = f64::from_bits(ui); |
| 106 | + |
| 107 | + let f: f64 = x - 1.0; |
| 108 | + let hfsq: f64 = 0.5 * f * f; |
| 109 | + let s: f64 = f / (2.0 + f); |
| 110 | + let z: f64 = s * s; |
| 111 | + let w: f64 = z * z; |
| 112 | + let t1: f64 = w * (LG2 + w * (LG4 + w * LG6)); |
| 113 | + let t2: f64 = z * (LG1 + w * (LG3 + w * (LG5 + w * LG7))); |
| 114 | + let r: f64 = t2 + t1; |
| 115 | + let dk: f64 = k as f64; |
| 116 | + return s * (hfsq + r) + dk * LN2_LO - hfsq + f + dk * LN2_HI; |
| 117 | +} |
0 commit comments