1 /* origin: FreeBSD /usr/src/lib/msun/src/s_cbrt.c */
2 /*
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  *
6  * Developed at SunPro, a Sun Microsystems, Inc. business.
7  * Permission to use, copy, modify, and distribute this
8  * software is freely granted, provided that this notice
9  * is preserved.
10  * ====================================================
11  *
12  * Optimized by Bruce D. Evans.
13  */
14 /* cbrt(x)
15  * Return cube root of x
16  */
17 
18 #include <math.h>
19 #include <stdint.h>
20 
21 static const uint32_t
22 B1 = 715094163, /* B1 = (1023-1023/3-0.03306235651)*2**20 */
23 B2 = 696219795; /* B2 = (1023-1023/3-54/3-0.03306235651)*2**20 */
24 
25 /* |1/cbrt(x) - p(x)| < 2**-23.5 (~[-7.93e-8, 7.929e-8]). */
26 static const double
27 P0 =  1.87595182427177009643,  /* 0x3ffe03e6, 0x0f61e692 */
28 P1 = -1.88497979543377169875,  /* 0xbffe28e0, 0x92f02420 */
29 P2 =  1.621429720105354466140, /* 0x3ff9f160, 0x4a49d6c2 */
30 P3 = -0.758397934778766047437, /* 0xbfe844cb, 0xbee751d9 */
31 P4 =  0.145996192886612446982; /* 0x3fc2b000, 0xd4e4edd7 */
32 
cbrt(double x)33 double cbrt(double x)
34 {
35 	union {double f; uint64_t i;} u = {x};
36 	double_t r,s,t,w;
37 	uint32_t hx = u.i>>32 & 0x7fffffff;
38 
39 	if (hx >= 0x7ff00000)  /* cbrt(NaN,INF) is itself */
40 		return x+x;
41 
42 	/*
43 	 * Rough cbrt to 5 bits:
44 	 *    cbrt(2**e*(1+m) ~= 2**(e/3)*(1+(e%3+m)/3)
45 	 * where e is integral and >= 0, m is real and in [0, 1), and "/" and
46 	 * "%" are integer division and modulus with rounding towards minus
47 	 * infinity.  The RHS is always >= the LHS and has a maximum relative
48 	 * error of about 1 in 16.  Adding a bias of -0.03306235651 to the
49 	 * (e%3+m)/3 term reduces the error to about 1 in 32. With the IEEE
50 	 * floating point representation, for finite positive normal values,
51 	 * ordinary integer divison of the value in bits magically gives
52 	 * almost exactly the RHS of the above provided we first subtract the
53 	 * exponent bias (1023 for doubles) and later add it back.  We do the
54 	 * subtraction virtually to keep e >= 0 so that ordinary integer
55 	 * division rounds towards minus infinity; this is also efficient.
56 	 */
57 	if (hx < 0x00100000) { /* zero or subnormal? */
58 		u.f = x*0x1p54;
59 		hx = u.i>>32 & 0x7fffffff;
60 		if (hx == 0)
61 			return x;  /* cbrt(0) is itself */
62 		hx = hx/3 + B2;
63 	} else
64 		hx = hx/3 + B1;
65 	u.i &= 1ULL<<63;
66 	u.i |= (uint64_t)hx << 32;
67 	t = u.f;
68 
69 	/*
70 	 * New cbrt to 23 bits:
71 	 *    cbrt(x) = t*cbrt(x/t**3) ~= t*P(t**3/x)
72 	 * where P(r) is a polynomial of degree 4 that approximates 1/cbrt(r)
73 	 * to within 2**-23.5 when |r - 1| < 1/10.  The rough approximation
74 	 * has produced t such than |t/cbrt(x) - 1| ~< 1/32, and cubing this
75 	 * gives us bounds for r = t**3/x.
76 	 *
77 	 * Try to optimize for parallel evaluation as in __tanf.c.
78 	 */
79 	r = (t*t)*(t/x);
80 	t = t*((P0+r*(P1+r*P2))+((r*r)*r)*(P3+r*P4));
81 
82 	/*
83 	 * Round t away from zero to 23 bits (sloppily except for ensuring that
84 	 * the result is larger in magnitude than cbrt(x) but not much more than
85 	 * 2 23-bit ulps larger).  With rounding towards zero, the error bound
86 	 * would be ~5/6 instead of ~4/6.  With a maximum error of 2 23-bit ulps
87 	 * in the rounded t, the infinite-precision error in the Newton
88 	 * approximation barely affects third digit in the final error
89 	 * 0.667; the error in the rounded t can be up to about 3 23-bit ulps
90 	 * before the final error is larger than 0.667 ulps.
91 	 */
92 	u.f = t;
93 	u.i = (u.i + 0x80000000) & 0xffffffffc0000000ULL;
94 	t = u.f;
95 
96 	/* one step Newton iteration to 53 bits with error < 0.667 ulps */
97 	s = t*t;         /* t*t is exact */
98 	r = x/s;         /* error <= 0.5 ulps; |r| < |t| */
99 	w = t+t;         /* t+t is exact */
100 	r = (r-t)/(w+r); /* r-t is exact; w+r ~= 3*t */
101 	t = t+t*r;       /* error <= 0.5 + 0.5/3 + epsilon */
102 	return t;
103 }
104