13a8617a8SJordan K. Hubbard /* k_tanf.c -- float version of k_tan.c 23a8617a8SJordan K. Hubbard * Conversion to float by Ian Lance Taylor, Cygnus Support, ian@cygnus.com. 3960d3da0SBruce Evans * Optimized by Bruce D. Evans. 43a8617a8SJordan K. Hubbard */ 53a8617a8SJordan K. Hubbard 63a8617a8SJordan K. Hubbard /* 73a8617a8SJordan K. Hubbard * ==================================================== 873fbb89dSDavid Schultz * Copyright 2004 Sun Microsystems, Inc. All Rights Reserved. 93a8617a8SJordan K. Hubbard * 103a8617a8SJordan K. Hubbard * Permission to use, copy, modify, and distribute this 113a8617a8SJordan K. Hubbard * software is freely granted, provided that this notice 123a8617a8SJordan K. Hubbard * is preserved. 133a8617a8SJordan K. Hubbard * ==================================================== 143a8617a8SJordan K. Hubbard */ 153a8617a8SJordan K. Hubbard 16960d3da0SBruce Evans #ifndef INLINE_KERNEL_TANDF 173a8617a8SJordan K. Hubbard #ifndef lint 187f3dea24SPeter Wemm static char rcsid[] = "$FreeBSD$"; 193a8617a8SJordan K. Hubbard #endif 204ce51209SBruce Evans #endif 213a8617a8SJordan K. Hubbard 223a8617a8SJordan K. Hubbard #include "math.h" 233a8617a8SJordan K. Hubbard #include "math_private.h" 24e96c4fd9SBruce Evans 2516638b55SBruce Evans /* |tan(x)/x - t(x)| < 2**-25.5 (~[-2e-08, 2e-08]). */ 2694a5f9beSBruce Evans static const double 273a8617a8SJordan K. Hubbard T[] = { 2816638b55SBruce Evans 0x15554d3418c99f.0p-54, /* 0.333331395030791399758 */ 2916638b55SBruce Evans 0x1112fd38999f72.0p-55, /* 0.133392002712976742718 */ 3016638b55SBruce Evans 0x1b54c91d865afe.0p-57, /* 0.0533812378445670393523 */ 3116638b55SBruce Evans 0x191df3908c33ce.0p-58, /* 0.0245283181166547278873 */ 3216638b55SBruce Evans 0x185dadfcecf44e.0p-61, /* 0.00297435743359967304927 */ 3316638b55SBruce Evans 0x1362b9bf971bcd.0p-59, /* 0.00946564784943673166728 */ 343a8617a8SJordan K. Hubbard }; 353a8617a8SJordan K. Hubbard 36960d3da0SBruce Evans #ifdef INLINE_KERNEL_TANDF 374ce51209SBruce Evans extern inline 384ce51209SBruce Evans #endif 3959b19ff1SAlfred Perlstein float 4094a5f9beSBruce Evans __kernel_tandf(double x, int iy) 413a8617a8SJordan K. Hubbard { 421dd21062SBruce Evans double z,r,w,s,t,u; 43e96c4fd9SBruce Evans 443a8617a8SJordan K. Hubbard z = x*x; 451dd21062SBruce Evans /* 461dd21062SBruce Evans * Split up the polynomial into small independent terms to give 471dd21062SBruce Evans * opportunities for parallel evaluation. The chosen splitting is 481dd21062SBruce Evans * micro-optimized for Athlons (XP, X64). It costs 2 multiplications 491dd21062SBruce Evans * relative to Horner's method on sequential machines. 501dd21062SBruce Evans * 511dd21062SBruce Evans * We add the small terms from lowest degree up for efficiency on 521dd21062SBruce Evans * non-sequential machines (the lowest degree terms tend to be ready 531dd21062SBruce Evans * earlier). Apart from this, we don't care about order of 541dd21062SBruce Evans * operations, and don't need to to care since we have precision to 551dd21062SBruce Evans * spare. However, the chosen splitting is good for accuracy too, 561dd21062SBruce Evans * and would give results as accurate as Horner's method if the 571dd21062SBruce Evans * small terms were added from highest degree down. 583a8617a8SJordan K. Hubbard */ 591dd21062SBruce Evans r = T[4]+z*T[5]; 601dd21062SBruce Evans t = T[2]+z*T[3]; 611dd21062SBruce Evans w = z*z; 623a8617a8SJordan K. Hubbard s = z*x; 631dd21062SBruce Evans u = T[0]+z*T[1]; 641dd21062SBruce Evans r = (x+s*u)+(s*w)*(t+w*r); 65833f0e1aSBruce Evans if(iy==1) return r; 66833f0e1aSBruce Evans else return -1.0/r; 673a8617a8SJordan K. Hubbard } 68