1 |
|
|
/* $OpenBSD: k_cosl.c,v 1.2 2017/01/21 08:29:13 krw Exp $ */ |
2 |
|
|
/* From: @(#)k_cos.c 1.3 95/01/18 */ |
3 |
|
|
/* |
4 |
|
|
* ==================================================== |
5 |
|
|
* Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
6 |
|
|
* Copyright (c) 2008 Steven G. Kargl, David Schultz, Bruce D. Evans. |
7 |
|
|
* |
8 |
|
|
* Developed at SunSoft, a Sun Microsystems, Inc. business. |
9 |
|
|
* Permission to use, copy, modify, and distribute this |
10 |
|
|
* software is freely granted, provided that this notice |
11 |
|
|
* is preserved. |
12 |
|
|
* ==================================================== |
13 |
|
|
*/ |
14 |
|
|
|
15 |
|
|
/* |
16 |
|
|
* ld80 version of k_cos.c. See ../k_cos.c for most comments. |
17 |
|
|
*/ |
18 |
|
|
|
19 |
|
|
#include "math_private.h" |
20 |
|
|
|
21 |
|
|
/* |
22 |
|
|
* Domain [-0.7854, 0.7854], range ~[-2.43e-23, 2.425e-23]: |
23 |
|
|
* |cos(x) - c(x)| < 2**-75.1 |
24 |
|
|
* |
25 |
|
|
* The coefficients of c(x) were generated by a pari-gp script using |
26 |
|
|
* a Remez algorithm that searches for the best higher coefficients |
27 |
|
|
* after rounding leading coefficients to a specified precision. |
28 |
|
|
* |
29 |
|
|
* Simpler methods like Chebyshev or basic Remez barely suffice for |
30 |
|
|
* cos() in 64-bit precision, because we want the coefficient of x^2 |
31 |
|
|
* to be precisely -0.5 so that multiplying by it is exact, and plain |
32 |
|
|
* rounding of the coefficients of a good polynomial approximation only |
33 |
|
|
* gives this up to about 64-bit precision. Plain rounding also gives |
34 |
|
|
* a mediocre approximation for the coefficient of x^4, but a rounding |
35 |
|
|
* error of 0.5 ulps for this coefficient would only contribute ~0.01 |
36 |
|
|
* ulps to the final error, so this is unimportant. Rounding errors in |
37 |
|
|
* higher coefficients are even less important. |
38 |
|
|
* |
39 |
|
|
* In fact, coefficients above the x^4 one only need to have 53-bit |
40 |
|
|
* precision, and this is more efficient. We get this optimization |
41 |
|
|
* almost for free from the complications needed to search for the best |
42 |
|
|
* higher coefficients. |
43 |
|
|
*/ |
44 |
|
|
static const double |
45 |
|
|
one = 1.0; |
46 |
|
|
|
47 |
|
|
#if defined(__amd64__) || defined(__i386__) |
48 |
|
|
/* Long double constants are slow on these arches, and broken on i386. */ |
49 |
|
|
static const volatile double |
50 |
|
|
C1hi = 0.041666666666666664, /* 0x15555555555555.0p-57 */ |
51 |
|
|
C1lo = 2.2598839032744733e-18; /* 0x14d80000000000.0p-111 */ |
52 |
|
|
#define C1 ((long double)C1hi + C1lo) |
53 |
|
|
#else |
54 |
|
|
static const long double |
55 |
|
|
C1 = 0.0416666666666666666136L; /* 0xaaaaaaaaaaaaaa9b.0p-68 */ |
56 |
|
|
#endif |
57 |
|
|
|
58 |
|
|
static const double |
59 |
|
|
C2 = -0.0013888888888888874, /* -0x16c16c16c16c10.0p-62 */ |
60 |
|
|
C3 = 0.000024801587301571716, /* 0x1a01a01a018e22.0p-68 */ |
61 |
|
|
C4 = -0.00000027557319215507120, /* -0x127e4fb7602f22.0p-74 */ |
62 |
|
|
C5 = 0.0000000020876754400407278, /* 0x11eed8caaeccf1.0p-81 */ |
63 |
|
|
C6 = -1.1470297442401303e-11, /* -0x19393412bd1529.0p-89 */ |
64 |
|
|
C7 = 4.7383039476436467e-14; /* 0x1aac9d9af5c43e.0p-97 */ |
65 |
|
|
|
66 |
|
|
long double |
67 |
|
|
__kernel_cosl(long double x, long double y) |
68 |
|
|
{ |
69 |
|
|
long double hz,z,r,w; |
70 |
|
|
|
71 |
|
392 |
z = x*x; |
72 |
|
196 |
r = z*(C1+z*(C2+z*(C3+z*(C4+z*(C5+z*(C6+z*C7)))))); |
73 |
|
196 |
hz = 0.5*z; |
74 |
|
196 |
w = one-hz; |
75 |
|
196 |
return w + (((one-w)-hz) + (z*r-x*y)); |
76 |
|
|
} |