1*1e651e1eSRoland Levillain 2*1e651e1eSRoland Levillain /* @(#)k_cos.c 1.3 95/01/18 */ 3*1e651e1eSRoland Levillain /* 4*1e651e1eSRoland Levillain * ==================================================== 5*1e651e1eSRoland Levillain * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. 6*1e651e1eSRoland Levillain * 7*1e651e1eSRoland Levillain * Developed at SunSoft, a Sun Microsystems, Inc. business. 8*1e651e1eSRoland Levillain * Permission to use, copy, modify, and distribute this 9*1e651e1eSRoland Levillain * software is freely granted, provided that this notice 10*1e651e1eSRoland Levillain * is preserved. 11*1e651e1eSRoland Levillain * ==================================================== 12*1e651e1eSRoland Levillain */ 13*1e651e1eSRoland Levillain 14*1e651e1eSRoland Levillain /* 15*1e651e1eSRoland Levillain * __kernel_cos( x, y ) 16*1e651e1eSRoland Levillain * kernel cos function on [-pi/4, pi/4], pi/4 ~ 0.785398164 17*1e651e1eSRoland Levillain * Input x is assumed to be bounded by ~pi/4 in magnitude. 18*1e651e1eSRoland Levillain * Input y is the tail of x. 19*1e651e1eSRoland Levillain * 20*1e651e1eSRoland Levillain * Algorithm 21*1e651e1eSRoland Levillain * 1. Since ieee_cos(-x) = ieee_cos(x), we need only to consider positive x. 22*1e651e1eSRoland Levillain * 2. if x < 2^-27 (hx<0x3e400000 0), return 1 with inexact if x!=0. 23*1e651e1eSRoland Levillain * 3. ieee_cos(x) is approximated by a polynomial of degree 14 on 24*1e651e1eSRoland Levillain * [0,pi/4] 25*1e651e1eSRoland Levillain * 4 14 26*1e651e1eSRoland Levillain * cos(x) ~ 1 - x*x/2 + C1*x + ... + C6*x 27*1e651e1eSRoland Levillain * where the remez error is 28*1e651e1eSRoland Levillain * 29*1e651e1eSRoland Levillain * | 2 4 6 8 10 12 14 | -58 30*1e651e1eSRoland Levillain * |ieee_cos(x)-(1-.5*x +C1*x +C2*x +C3*x +C4*x +C5*x +C6*x )| <= 2 31*1e651e1eSRoland Levillain * | | 32*1e651e1eSRoland Levillain * 33*1e651e1eSRoland Levillain * 4 6 8 10 12 14 34*1e651e1eSRoland Levillain * 4. let r = C1*x +C2*x +C3*x +C4*x +C5*x +C6*x , then 35*1e651e1eSRoland Levillain * ieee_cos(x) = 1 - x*x/2 + r 36*1e651e1eSRoland Levillain * since ieee_cos(x+y) ~ ieee_cos(x) - ieee_sin(x)*y 37*1e651e1eSRoland Levillain * ~ ieee_cos(x) - x*y, 38*1e651e1eSRoland Levillain * a correction term is necessary in ieee_cos(x) and hence 39*1e651e1eSRoland Levillain * cos(x+y) = 1 - (x*x/2 - (r - x*y)) 40*1e651e1eSRoland Levillain * For better accuracy when x > 0.3, let qx = |x|/4 with 41*1e651e1eSRoland Levillain * the last 32 bits mask off, and if x > 0.78125, let qx = 0.28125. 42*1e651e1eSRoland Levillain * Then 43*1e651e1eSRoland Levillain * cos(x+y) = (1-qx) - ((x*x/2-qx) - (r-x*y)). 44*1e651e1eSRoland Levillain * Note that 1-qx and (x*x/2-qx) is EXACT here, and the 45*1e651e1eSRoland Levillain * magnitude of the latter is at least a quarter of x*x/2, 46*1e651e1eSRoland Levillain * thus, reducing the rounding error in the subtraction. 47*1e651e1eSRoland Levillain */ 48*1e651e1eSRoland Levillain 49*1e651e1eSRoland Levillain #include "fdlibm.h" 50*1e651e1eSRoland Levillain 51*1e651e1eSRoland Levillain #ifdef __STDC__ 52*1e651e1eSRoland Levillain static const double 53*1e651e1eSRoland Levillain #else 54*1e651e1eSRoland Levillain static double 55*1e651e1eSRoland Levillain #endif 56*1e651e1eSRoland Levillain one = 1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */ 57*1e651e1eSRoland Levillain C1 = 4.16666666666666019037e-02, /* 0x3FA55555, 0x5555554C */ 58*1e651e1eSRoland Levillain C2 = -1.38888888888741095749e-03, /* 0xBF56C16C, 0x16C15177 */ 59*1e651e1eSRoland Levillain C3 = 2.48015872894767294178e-05, /* 0x3EFA01A0, 0x19CB1590 */ 60*1e651e1eSRoland Levillain C4 = -2.75573143513906633035e-07, /* 0xBE927E4F, 0x809C52AD */ 61*1e651e1eSRoland Levillain C5 = 2.08757232129817482790e-09, /* 0x3E21EE9E, 0xBDB4B1C4 */ 62*1e651e1eSRoland Levillain C6 = -1.13596475577881948265e-11; /* 0xBDA8FAE9, 0xBE8838D4 */ 63*1e651e1eSRoland Levillain 64*1e651e1eSRoland Levillain #ifdef __STDC__ __kernel_cos(double x,double y)65*1e651e1eSRoland Levillain double __kernel_cos(double x, double y) 66*1e651e1eSRoland Levillain #else 67*1e651e1eSRoland Levillain double __kernel_cos(x, y) 68*1e651e1eSRoland Levillain double x,y; 69*1e651e1eSRoland Levillain #endif 70*1e651e1eSRoland Levillain { 71*1e651e1eSRoland Levillain double a,hz,z,r,qx; 72*1e651e1eSRoland Levillain int ix; 73*1e651e1eSRoland Levillain ix = __HI(x)&0x7fffffff; /* ix = |x|'s high word*/ 74*1e651e1eSRoland Levillain if(ix<0x3e400000) { /* if x < 2**27 */ 75*1e651e1eSRoland Levillain if(((int)x)==0) return one; /* generate inexact */ 76*1e651e1eSRoland Levillain } 77*1e651e1eSRoland Levillain z = x*x; 78*1e651e1eSRoland Levillain r = z*(C1+z*(C2+z*(C3+z*(C4+z*(C5+z*C6))))); 79*1e651e1eSRoland Levillain if(ix < 0x3FD33333) /* if |x| < 0.3 */ 80*1e651e1eSRoland Levillain return one - (0.5*z - (z*r - x*y)); 81*1e651e1eSRoland Levillain else { 82*1e651e1eSRoland Levillain if(ix > 0x3fe90000) { /* x > 0.78125 */ 83*1e651e1eSRoland Levillain qx = 0.28125; 84*1e651e1eSRoland Levillain } else { 85*1e651e1eSRoland Levillain __HI(qx) = ix-0x00200000; /* x/4 */ 86*1e651e1eSRoland Levillain __LO(qx) = 0; 87*1e651e1eSRoland Levillain } 88*1e651e1eSRoland Levillain hz = 0.5*z-qx; 89*1e651e1eSRoland Levillain a = one-qx; 90*1e651e1eSRoland Levillain return a - (hz - (z*r-x*y)); 91*1e651e1eSRoland Levillain } 92*1e651e1eSRoland Levillain } 93