OLD | NEW |
1 /* origin: FreeBSD /usr/src/lib/msun/src/s_cbrt.c */ | 1 /* origin: FreeBSD /usr/src/lib/msun/src/s_cbrt.c */ |
2 /* | 2 /* |
3 * ==================================================== | 3 * ==================================================== |
4 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. | 4 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
5 * | 5 * |
6 * Developed at SunPro, a Sun Microsystems, Inc. business. | 6 * Developed at SunPro, a Sun Microsystems, Inc. business. |
7 * Permission to use, copy, modify, and distribute this | 7 * Permission to use, copy, modify, and distribute this |
8 * software is freely granted, provided that this notice | 8 * software is freely granted, provided that this notice |
9 * is preserved. | 9 * is preserved. |
10 * ==================================================== | 10 * ==================================================== |
11 * | 11 * |
12 * Optimized by Bruce D. Evans. | 12 * Optimized by Bruce D. Evans. |
13 */ | 13 */ |
14 /* cbrt(x) | 14 /* cbrt(x) |
15 * Return cube root of x | 15 * Return cube root of x |
16 */ | 16 */ |
17 | 17 |
18 #include <math.h> | 18 #include <math.h> |
19 #include <stdint.h> | 19 #include <stdint.h> |
20 | 20 |
21 static const uint32_t | 21 static const uint32_t |
22 B1 = 715094163, /* B1 = (1023-1023/3-0.03306235651)*2**20 */ | 22 B1 = 715094163, /* B1 = (1023-1023/3-0.03306235651)*2**20 */ |
23 B2 = 696219795; /* B2 = (1023-1023/3-54/3-0.03306235651)*2**20 */ | 23 B2 = 696219795; /* B2 = (1023-1023/3-54/3-0.03306235651)*2**20 */ |
24 | 24 |
25 /* |1/cbrt(x) - p(x)| < 2**-23.5 (~[-7.93e-8, 7.929e-8]). */ | 25 /* |1/cbrt(x) - p(x)| < 2**-23.5 (~[-7.93e-8, 7.929e-8]). */ |
26 static const double | 26 static const double P0 = 1.87595182427177009643, /* 0x3ffe03e6, 0x0f61e692 */ |
27 P0 = 1.87595182427177009643, /* 0x3ffe03e6, 0x0f61e692 */ | 27 P1 = -1.88497979543377169875, /* 0xbffe28e0, 0x92f02420 */ |
28 P1 = -1.88497979543377169875, /* 0xbffe28e0, 0x92f02420 */ | 28 P2 = 1.621429720105354466140, /* 0x3ff9f160, 0x4a49d6c2 */ |
29 P2 = 1.621429720105354466140, /* 0x3ff9f160, 0x4a49d6c2 */ | 29 P3 = -0.758397934778766047437, /* 0xbfe844cb, 0xbee751d9 */ |
30 P3 = -0.758397934778766047437, /* 0xbfe844cb, 0xbee751d9 */ | 30 P4 = 0.145996192886612446982; /* 0x3fc2b000, 0xd4e4edd7 */ |
31 P4 = 0.145996192886612446982; /* 0x3fc2b000, 0xd4e4edd7 */ | |
32 | 31 |
33 double cbrt(double x) | 32 double cbrt(double x) { |
34 { | 33 union { |
35 » union {double f; uint64_t i;} u = {x}; | 34 double f; |
36 » double_t r,s,t,w; | 35 uint64_t i; |
37 » uint32_t hx = u.i>>32 & 0x7fffffff; | 36 } u = {x}; |
| 37 double_t r, s, t, w; |
| 38 uint32_t hx = u.i >> 32 & 0x7fffffff; |
38 | 39 |
39 » if (hx >= 0x7ff00000) /* cbrt(NaN,INF) is itself */ | 40 if (hx >= 0x7ff00000) /* cbrt(NaN,INF) is itself */ |
40 » » return x+x; | 41 return x + x; |
41 | 42 |
42 » /* | 43 /* |
43 » * Rough cbrt to 5 bits: | 44 * Rough cbrt to 5 bits: |
44 » * cbrt(2**e*(1+m) ~= 2**(e/3)*(1+(e%3+m)/3) | 45 * 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 * 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 * "%" are integer division and modulus with rounding towards minus |
47 » * infinity. The RHS is always >= the LHS and has a maximum relative | 48 * 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 * 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 * (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 * floating point representation, for finite positive normal values, |
51 » * ordinary integer divison of the value in bits magically gives | 52 * ordinary integer divison of the value in bits magically gives |
52 » * almost exactly the RHS of the above provided we first subtract the | 53 * 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 * 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 * subtraction virtually to keep e >= 0 so that ordinary integer |
55 » * division rounds towards minus infinity; this is also efficient. | 56 * division rounds towards minus infinity; this is also efficient. |
56 » */ | 57 */ |
57 » if (hx < 0x00100000) { /* zero or subnormal? */ | 58 if (hx < 0x00100000) { /* zero or subnormal? */ |
58 » » u.f = x*0x1p54; | 59 u.f = x * 0x1p54; |
59 » » hx = u.i>>32 & 0x7fffffff; | 60 hx = u.i >> 32 & 0x7fffffff; |
60 » » if (hx == 0) | 61 if (hx == 0) |
61 » » » return x; /* cbrt(0) is itself */ | 62 return x; /* cbrt(0) is itself */ |
62 » » hx = hx/3 + B2; | 63 hx = hx / 3 + B2; |
63 » } else | 64 } else |
64 » » hx = hx/3 + B1; | 65 hx = hx / 3 + B1; |
65 » u.i &= 1ULL<<63; | 66 u.i &= 1ULL << 63; |
66 » u.i |= (uint64_t)hx << 32; | 67 u.i |= (uint64_t)hx << 32; |
67 » t = u.f; | 68 t = u.f; |
68 | 69 |
69 » /* | 70 /* |
70 » * New cbrt to 23 bits: | 71 * New cbrt to 23 bits: |
71 » * cbrt(x) = t*cbrt(x/t**3) ~= t*P(t**3/x) | 72 * 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 * 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 * 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 * has produced t such than |t/cbrt(x) - 1| ~< 1/32, and cubing this |
75 » * gives us bounds for r = t**3/x. | 76 * gives us bounds for r = t**3/x. |
76 » * | 77 * |
77 » * Try to optimize for parallel evaluation as in __tanf.c. | 78 * Try to optimize for parallel evaluation as in __tanf.c. |
78 » */ | 79 */ |
79 » r = (t*t)*(t/x); | 80 r = (t * t) * (t / x); |
80 » t = t*((P0+r*(P1+r*P2))+((r*r)*r)*(P3+r*P4)); | 81 t = t * ((P0 + r * (P1 + r * P2)) + ((r * r) * r) * (P3 + r * P4)); |
81 | 82 |
82 » /* | 83 /* |
83 » * Round t away from zero to 23 bits (sloppily except for ensuring that | 84 * 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 * 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 * 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 * 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 * in the rounded t, the infinite-precision error in the Newton |
88 » * approximation barely affects third digit in the final error | 89 * 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 * 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 * before the final error is larger than 0.667 ulps. |
91 » */ | 92 */ |
92 » u.f = t; | 93 u.f = t; |
93 » u.i = (u.i + 0x80000000) & 0xffffffffc0000000ULL; | 94 u.i = (u.i + 0x80000000) & 0xffffffffc0000000ULL; |
94 » t = u.f; | 95 t = u.f; |
95 | 96 |
96 » /* one step Newton iteration to 53 bits with error < 0.667 ulps */ | 97 /* one step Newton iteration to 53 bits with error < 0.667 ulps */ |
97 » s = t*t; /* t*t is exact */ | 98 s = t * t; /* t*t is exact */ |
98 » r = x/s; /* error <= 0.5 ulps; |r| < |t| */ | 99 r = x / s; /* error <= 0.5 ulps; |r| < |t| */ |
99 » w = t+t; /* t+t is exact */ | 100 w = t + t; /* t+t is exact */ |
100 » r = (r-t)/(w+r); /* r-t is exact; w+r ~= 3*t */ | 101 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 t = t + t * r; /* error <= 0.5 + 0.5/3 + epsilon */ |
102 » return t; | 103 return t; |
103 } | 104 } |
OLD | NEW |