13a8617a8SJordan K. Hubbard /* @(#)k_cos.c 5.1 93/09/24 */ 23a8617a8SJordan K. Hubbard /* 33a8617a8SJordan K. Hubbard * ==================================================== 43a8617a8SJordan K. Hubbard * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. 53a8617a8SJordan K. Hubbard * 63a8617a8SJordan K. Hubbard * Developed at SunPro, 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 #ifndef lint 143a8617a8SJordan K. Hubbard static char rcsid[] = "$Id: k_cos.c,v 1.6 1994/08/18 23:06:08 jtc Exp $"; 153a8617a8SJordan K. Hubbard #endif 163a8617a8SJordan K. Hubbard 173a8617a8SJordan K. Hubbard /* 183a8617a8SJordan K. Hubbard * __kernel_cos( x, y ) 193a8617a8SJordan K. Hubbard * kernel cos function on [-pi/4, pi/4], pi/4 ~ 0.785398164 203a8617a8SJordan K. Hubbard * Input x is assumed to be bounded by ~pi/4 in magnitude. 213a8617a8SJordan K. Hubbard * Input y is the tail of x. 223a8617a8SJordan K. Hubbard * 233a8617a8SJordan K. Hubbard * Algorithm 243a8617a8SJordan K. Hubbard * 1. Since cos(-x) = cos(x), we need only to consider positive x. 253a8617a8SJordan K. Hubbard * 2. if x < 2^-27 (hx<0x3e400000 0), return 1 with inexact if x!=0. 263a8617a8SJordan K. Hubbard * 3. cos(x) is approximated by a polynomial of degree 14 on 273a8617a8SJordan K. Hubbard * [0,pi/4] 283a8617a8SJordan K. Hubbard * 4 14 293a8617a8SJordan K. Hubbard * cos(x) ~ 1 - x*x/2 + C1*x + ... + C6*x 303a8617a8SJordan K. Hubbard * where the remez error is 313a8617a8SJordan K. Hubbard * 323a8617a8SJordan K. Hubbard * | 2 4 6 8 10 12 14 | -58 333a8617a8SJordan K. Hubbard * |cos(x)-(1-.5*x +C1*x +C2*x +C3*x +C4*x +C5*x +C6*x )| <= 2 343a8617a8SJordan K. Hubbard * | | 353a8617a8SJordan K. Hubbard * 363a8617a8SJordan K. Hubbard * 4 6 8 10 12 14 373a8617a8SJordan K. Hubbard * 4. let r = C1*x +C2*x +C3*x +C4*x +C5*x +C6*x , then 383a8617a8SJordan K. Hubbard * cos(x) = 1 - x*x/2 + r 393a8617a8SJordan K. Hubbard * since cos(x+y) ~ cos(x) - sin(x)*y 403a8617a8SJordan K. Hubbard * ~ cos(x) - x*y, 413a8617a8SJordan K. Hubbard * a correction term is necessary in cos(x) and hence 423a8617a8SJordan K. Hubbard * cos(x+y) = 1 - (x*x/2 - (r - x*y)) 433a8617a8SJordan K. Hubbard * For better accuracy when x > 0.3, let qx = |x|/4 with 443a8617a8SJordan K. Hubbard * the last 32 bits mask off, and if x > 0.78125, let qx = 0.28125. 453a8617a8SJordan K. Hubbard * Then 463a8617a8SJordan K. Hubbard * cos(x+y) = (1-qx) - ((x*x/2-qx) - (r-x*y)). 473a8617a8SJordan K. Hubbard * Note that 1-qx and (x*x/2-qx) is EXACT here, and the 483a8617a8SJordan K. Hubbard * magnitude of the latter is at least a quarter of x*x/2, 493a8617a8SJordan K. Hubbard * thus, reducing the rounding error in the subtraction. 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 #ifdef __STDC__ 563a8617a8SJordan K. Hubbard static const double 573a8617a8SJordan K. Hubbard #else 583a8617a8SJordan K. Hubbard static double 593a8617a8SJordan K. Hubbard #endif 603a8617a8SJordan K. Hubbard one = 1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */ 613a8617a8SJordan K. Hubbard C1 = 4.16666666666666019037e-02, /* 0x3FA55555, 0x5555554C */ 623a8617a8SJordan K. Hubbard C2 = -1.38888888888741095749e-03, /* 0xBF56C16C, 0x16C15177 */ 633a8617a8SJordan K. Hubbard C3 = 2.48015872894767294178e-05, /* 0x3EFA01A0, 0x19CB1590 */ 643a8617a8SJordan K. Hubbard C4 = -2.75573143513906633035e-07, /* 0xBE927E4F, 0x809C52AD */ 653a8617a8SJordan K. Hubbard C5 = 2.08757232129817482790e-09, /* 0x3E21EE9E, 0xBDB4B1C4 */ 663a8617a8SJordan K. Hubbard C6 = -1.13596475577881948265e-11; /* 0xBDA8FAE9, 0xBE8838D4 */ 673a8617a8SJordan K. Hubbard 683a8617a8SJordan K. Hubbard #ifdef __STDC__ 693a8617a8SJordan K. Hubbard double __kernel_cos(double x, double y) 703a8617a8SJordan K. Hubbard #else 713a8617a8SJordan K. Hubbard double __kernel_cos(x, y) 723a8617a8SJordan K. Hubbard double x,y; 733a8617a8SJordan K. Hubbard #endif 743a8617a8SJordan K. Hubbard { 753a8617a8SJordan K. Hubbard double a,hz,z,r,qx; 763a8617a8SJordan K. Hubbard int32_t ix; 773a8617a8SJordan K. Hubbard GET_HIGH_WORD(ix,x); 783a8617a8SJordan K. Hubbard ix &= 0x7fffffff; /* ix = |x|'s high word*/ 793a8617a8SJordan K. Hubbard if(ix<0x3e400000) { /* if x < 2**27 */ 803a8617a8SJordan K. Hubbard if(((int)x)==0) return one; /* generate inexact */ 813a8617a8SJordan K. Hubbard } 823a8617a8SJordan K. Hubbard z = x*x; 833a8617a8SJordan K. Hubbard r = z*(C1+z*(C2+z*(C3+z*(C4+z*(C5+z*C6))))); 843a8617a8SJordan K. Hubbard if(ix < 0x3FD33333) /* if |x| < 0.3 */ 853a8617a8SJordan K. Hubbard return one - (0.5*z - (z*r - x*y)); 863a8617a8SJordan K. Hubbard else { 873a8617a8SJordan K. Hubbard if(ix > 0x3fe90000) { /* x > 0.78125 */ 883a8617a8SJordan K. Hubbard qx = 0.28125; 893a8617a8SJordan K. Hubbard } else { 903a8617a8SJordan K. Hubbard INSERT_WORDS(qx,ix-0x00200000,0); /* x/4 */ 913a8617a8SJordan K. Hubbard } 923a8617a8SJordan K. Hubbard hz = 0.5*z-qx; 933a8617a8SJordan K. Hubbard a = one-qx; 943a8617a8SJordan K. Hubbard return a - (hz - (z*r-x*y)); 953a8617a8SJordan K. Hubbard } 963a8617a8SJordan K. Hubbard } 97