13f708241SDavid Schultz
23a8617a8SJordan K. Hubbard /*
33a8617a8SJordan K. Hubbard * ====================================================
43a8617a8SJordan K. Hubbard * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
53a8617a8SJordan K. Hubbard *
63f708241SDavid Schultz * Developed at SunSoft, a Sun Microsystems, Inc. business.
73a8617a8SJordan K. Hubbard * Permission to use, copy, modify, and distribute this
83a8617a8SJordan K. Hubbard * software is freely granted, provided that this notice
93a8617a8SJordan K. Hubbard * is preserved.
103a8617a8SJordan K. Hubbard * ====================================================
113a8617a8SJordan K. Hubbard */
123a8617a8SJordan K. Hubbard
133a8617a8SJordan K. Hubbard /*
143a8617a8SJordan K. Hubbard * __kernel_cos( x, y )
153a8617a8SJordan K. Hubbard * kernel cos function on [-pi/4, pi/4], pi/4 ~ 0.785398164
163a8617a8SJordan K. Hubbard * Input x is assumed to be bounded by ~pi/4 in magnitude.
173a8617a8SJordan K. Hubbard * Input y is the tail of x.
183a8617a8SJordan K. Hubbard *
193a8617a8SJordan K. Hubbard * Algorithm
203a8617a8SJordan K. Hubbard * 1. Since cos(-x) = cos(x), we need only to consider positive x.
213a8617a8SJordan K. Hubbard * 2. if x < 2^-27 (hx<0x3e400000 0), return 1 with inexact if x!=0.
223a8617a8SJordan K. Hubbard * 3. cos(x) is approximated by a polynomial of degree 14 on
233a8617a8SJordan K. Hubbard * [0,pi/4]
243a8617a8SJordan K. Hubbard * 4 14
253a8617a8SJordan K. Hubbard * cos(x) ~ 1 - x*x/2 + C1*x + ... + C6*x
263a8617a8SJordan K. Hubbard * where the remez error is
273a8617a8SJordan K. Hubbard *
283a8617a8SJordan K. Hubbard * | 2 4 6 8 10 12 14 | -58
293a8617a8SJordan K. Hubbard * |cos(x)-(1-.5*x +C1*x +C2*x +C3*x +C4*x +C5*x +C6*x )| <= 2
303a8617a8SJordan K. Hubbard * | |
313a8617a8SJordan K. Hubbard *
323a8617a8SJordan K. Hubbard * 4 6 8 10 12 14
333a8617a8SJordan K. Hubbard * 4. let r = C1*x +C2*x +C3*x +C4*x +C5*x +C6*x , then
343b46e988SBruce Evans * cos(x) ~ 1 - x*x/2 + r
353a8617a8SJordan K. Hubbard * since cos(x+y) ~ cos(x) - sin(x)*y
363a8617a8SJordan K. Hubbard * ~ cos(x) - x*y,
373a8617a8SJordan K. Hubbard * a correction term is necessary in cos(x) and hence
383a8617a8SJordan K. Hubbard * cos(x+y) = 1 - (x*x/2 - (r - x*y))
393b46e988SBruce Evans * For better accuracy, rearrange to
403b46e988SBruce Evans * cos(x+y) ~ w + (tmp + (r-x*y))
413b46e988SBruce Evans * where w = 1 - x*x/2 and tmp is a tiny correction term
423b46e988SBruce Evans * (1 - x*x/2 == w + tmp exactly in infinite precision).
433b46e988SBruce Evans * The exactness of w + tmp in infinite precision depends on w
443b46e988SBruce Evans * and tmp having the same precision as x. If they have extra
453b46e988SBruce Evans * precision due to compiler bugs, then the extra precision is
463b46e988SBruce Evans * only good provided it is retained in all terms of the final
473b46e988SBruce Evans * expression for cos(). Retention happens in all cases tested
483b46e988SBruce Evans * under FreeBSD, so don't pessimize things by forcibly clipping
493b46e988SBruce Evans * any extra precision in w.
503a8617a8SJordan K. Hubbard */
513a8617a8SJordan K. Hubbard
523a8617a8SJordan K. Hubbard #include "math.h"
533a8617a8SJordan K. Hubbard #include "math_private.h"
543a8617a8SJordan K. Hubbard
553a8617a8SJordan K. Hubbard static const double
563a8617a8SJordan K. Hubbard one = 1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */
573a8617a8SJordan K. Hubbard C1 = 4.16666666666666019037e-02, /* 0x3FA55555, 0x5555554C */
583a8617a8SJordan K. Hubbard C2 = -1.38888888888741095749e-03, /* 0xBF56C16C, 0x16C15177 */
593a8617a8SJordan K. Hubbard C3 = 2.48015872894767294178e-05, /* 0x3EFA01A0, 0x19CB1590 */
603a8617a8SJordan K. Hubbard C4 = -2.75573143513906633035e-07, /* 0xBE927E4F, 0x809C52AD */
613a8617a8SJordan K. Hubbard C5 = 2.08757232129817482790e-09, /* 0x3E21EE9E, 0xBDB4B1C4 */
623a8617a8SJordan K. Hubbard C6 = -1.13596475577881948265e-11; /* 0xBDA8FAE9, 0xBE8838D4 */
633a8617a8SJordan K. Hubbard
6459b19ff1SAlfred Perlstein double
__kernel_cos(double x,double y)6559b19ff1SAlfred Perlstein __kernel_cos(double x, double y)
663a8617a8SJordan K. Hubbard {
673b46e988SBruce Evans double hz,z,r,w;
683b46e988SBruce Evans
693a8617a8SJordan K. Hubbard z = x*x;
709ce87560SBruce Evans w = z*z;
719ce87560SBruce Evans r = z*(C1+z*(C2+z*C3)) + w*w*(C4+z*(C5+z*C6));
72fc84b771SBruce Evans hz = 0.5*z;
733b46e988SBruce Evans w = one-hz;
743b46e988SBruce Evans return w + (((one-w)-hz) + (z*r-x*y));
753a8617a8SJordan K. Hubbard }
76