13a8617a8SJordan K. Hubbard /* s_cbrtf.c -- float version of s_cbrt.c. 23a8617a8SJordan K. Hubbard * Conversion to float by Ian Lance Taylor, Cygnus Support, ian@cygnus.com. 3ec761d75SBruce Evans * Debugged and optimized by Bruce D. Evans. 43a8617a8SJordan K. Hubbard */ 53a8617a8SJordan K. Hubbard 63a8617a8SJordan K. Hubbard /* 73a8617a8SJordan K. Hubbard * ==================================================== 83a8617a8SJordan K. Hubbard * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. 93a8617a8SJordan K. Hubbard * 103a8617a8SJordan K. Hubbard * Developed at SunPro, a Sun Microsystems, Inc. business. 113a8617a8SJordan K. Hubbard * Permission to use, copy, modify, and distribute this 123a8617a8SJordan K. Hubbard * software is freely granted, provided that this notice 133a8617a8SJordan K. Hubbard * is preserved. 143a8617a8SJordan K. Hubbard * ==================================================== 153a8617a8SJordan K. Hubbard */ 163a8617a8SJordan K. Hubbard 173a8617a8SJordan K. Hubbard #ifndef lint 187f3dea24SPeter Wemm static char rcsid[] = "$FreeBSD$"; 193a8617a8SJordan K. Hubbard #endif 203a8617a8SJordan K. Hubbard 213a8617a8SJordan K. Hubbard #include "math.h" 223a8617a8SJordan K. Hubbard #include "math_private.h" 233a8617a8SJordan K. Hubbard 243a8617a8SJordan K. Hubbard /* cbrtf(x) 253a8617a8SJordan K. Hubbard * Return cube root of x 263a8617a8SJordan K. Hubbard */ 273a8617a8SJordan K. Hubbard static const unsigned 28af7f9913SBruce Evans B1 = 709958130, /* B1 = (127-127.0/3-0.03306235651)*2**23 */ 29af7f9913SBruce Evans B2 = 642849266; /* B2 = (127-127.0/3-24/3-0.03306235651)*2**23 */ 303a8617a8SJordan K. Hubbard 3159b19ff1SAlfred Perlstein float 3259b19ff1SAlfred Perlstein cbrtf(float x) 333a8617a8SJordan K. Hubbard { 344bb97803SBruce Evans double r,T; 354bb97803SBruce Evans float t; 363a8617a8SJordan K. Hubbard int32_t hx; 373a8617a8SJordan K. Hubbard u_int32_t sign; 383a8617a8SJordan K. Hubbard u_int32_t high; 393a8617a8SJordan K. Hubbard 403a8617a8SJordan K. Hubbard GET_FLOAT_WORD(hx,x); 413a8617a8SJordan K. Hubbard sign=hx&0x80000000; /* sign= sign(x) */ 423a8617a8SJordan K. Hubbard hx ^=sign; 433a8617a8SJordan K. Hubbard if(hx>=0x7f800000) return(x+x); /* cbrt(NaN,INF) is itself */ 443a8617a8SJordan K. Hubbard 453a8617a8SJordan K. Hubbard /* rough cbrt to 5 bits */ 4620a99011SBruce Evans if(hx<0x00800000) { /* zero or subnormal? */ 4720a99011SBruce Evans if(hx==0) 4820a99011SBruce Evans return(x); /* cbrt(+-0) is itself */ 497d5a4821SBruce Evans SET_FLOAT_WORD(t,0x4b800000); /* set t= 2**24 */ 507d5a4821SBruce Evans t*=x; 517d5a4821SBruce Evans GET_FLOAT_WORD(high,t); 52ec761d75SBruce Evans SET_FLOAT_WORD(t,sign|((high&0x7fffffff)/3+B2)); 537d5a4821SBruce Evans } else 54ec761d75SBruce Evans SET_FLOAT_WORD(t,sign|(hx/3+B1)); 553a8617a8SJordan K. Hubbard 56fd289100SBruce Evans /* 57fd289100SBruce Evans * First step Newton iteration (solving t*t-x/t == 0) to 16 bits. In 58fd289100SBruce Evans * double precision so that its terms can be arranged for efficiency 59fd289100SBruce Evans * without causing overflow or underflow. 60fd289100SBruce Evans */ 614bb97803SBruce Evans T=t; 624bb97803SBruce Evans r=T*T*T; 63fd289100SBruce Evans T=T*((double)x+x+r)/(x+r+r); 643a8617a8SJordan K. Hubbard 65fd289100SBruce Evans /* 66fd289100SBruce Evans * Second step Newton iteration to 47 bits. In double precision for 67fd289100SBruce Evans * efficiency and accuracy. 68fd289100SBruce Evans */ 694bb97803SBruce Evans r=T*T*T; 70fd289100SBruce Evans T=T*((double)x+x+r)/(x+r+r); 716de073b4SBruce Evans 724bb97803SBruce Evans /* rounding to 24 bits is perfect in round-to-nearest mode */ 734bb97803SBruce Evans return(T); 743a8617a8SJordan K. Hubbard } 75