1 /* $OpenBSD: e_powl.c,v 1.5 2013/11/12 20:35:19 martynas Exp $ */
4 * Copyright (c) 2008 Stephen L. Moshier <steve@moshier.net>
6 * Permission to use, copy, modify, and distribute this software for any
7 * purpose with or without fee is hereby granted, provided that the above
8 * copyright notice and this permission notice appear in all copies.
10 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
11 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
12 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
13 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
14 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
15 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
16 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
21 * Power function, long double precision
27 * long double x, y, z, powl();
35 * Computes x raised to the yth power. Analytically,
37 * x**y = exp( y log(x) ).
39 * Following Cody and Waite, this program uses a lookup table
40 * of 2**-i/32 and pseudo extended precision arithmetic to
41 * obtain several extra bits of accuracy in both the logarithm
42 * and the exponential.
48 * The relative error of pow(x,y) can be estimated
49 * by y dl ln(2), where dl is the absolute error of
50 * the internally computed base 2 logarithm. At the ends
51 * of the approximation interval the logarithm equal 1/32
52 * and its relative error is about 1 lsb = 1.1e-19. Hence
53 * the predicted relative error in the result is 2.3e-21 y .
56 * arithmetic domain # trials peak rms
58 * IEEE +-1000 40000 2.8e-18 3.7e-19
59 * .001 < x < 1000, with log(x) uniformly distributed.
60 * -1000 < y < 1000, y uniformly distributed.
62 * IEEE 0,8700 60000 6.5e-18 1.0e-18
63 * 0.99 < x < 1.01, 0 < y < 8700, uniformly distributed.
68 * message condition value returned
69 * pow overflow x**y > MAXNUM INFINITY
70 * pow underflow x**y < 1/MAXNUM 0.0
71 * pow domain x<0 and y noninteger 0.0
78 #include "math_private.h"
82 /* log2(Table size) */
85 /* log(1+x) = x - .5x^2 + x^3 * P(z)/Q(z)
86 * on the domain 2^(-1/32) - 1 <= x <= 2^(1/32) - 1
88 static const long double P
[] = {
89 8.3319510773868690346226E-4L,
90 4.9000050881978028599627E-1L,
91 1.7500123722550302671919E0L
,
92 1.4000100839971580279335E0L
,
94 static const long double Q
[] = {
95 /* 1.0000000000000000000000E0L,*/
96 5.2500282295834889175431E0L
,
97 8.4000598057587009834666E0L
,
98 4.2000302519914740834728E0L
,
100 /* A[i] = 2^(-i/32), rounded to IEEE long double precision.
101 * If i is even, A[i] + B[i/2] gives additional accuracy.
103 static const long double A
[33] = {
104 1.0000000000000000000000E0L
,
105 9.7857206208770013448287E-1L,
106 9.5760328069857364691013E-1L,
107 9.3708381705514995065011E-1L,
108 9.1700404320467123175367E-1L,
109 8.9735453750155359320742E-1L,
110 8.7812608018664974155474E-1L,
111 8.5930964906123895780165E-1L,
112 8.4089641525371454301892E-1L,
113 8.2287773907698242225554E-1L,
114 8.0524516597462715409607E-1L,
115 7.8799042255394324325455E-1L,
116 7.7110541270397041179298E-1L,
117 7.5458221379671136985669E-1L,
118 7.3841307296974965571198E-1L,
119 7.2259040348852331001267E-1L,
120 7.0710678118654752438189E-1L,
121 6.9195494098191597746178E-1L,
122 6.7712777346844636413344E-1L,
123 6.6261832157987064729696E-1L,
124 6.4841977732550483296079E-1L,
125 6.3452547859586661129850E-1L,
126 6.2092890603674202431705E-1L,
127 6.0762367999023443907803E-1L,
128 5.9460355750136053334378E-1L,
129 5.8186242938878875689693E-1L,
130 5.6939431737834582684856E-1L,
131 5.5719337129794626814472E-1L,
132 5.4525386633262882960438E-1L,
133 5.3357020033841180906486E-1L,
134 5.2213689121370692017331E-1L,
135 5.1094857432705833910408E-1L,
136 5.0000000000000000000000E-1L,
138 static const long double B
[17] = {
139 0.0000000000000000000000E0L
,
140 2.6176170809902549338711E-20L,
141 -1.0126791927256478897086E-20L,
142 1.3438228172316276937655E-21L,
143 1.2207982955417546912101E-20L,
144 -6.3084814358060867200133E-21L,
145 1.3164426894366316434230E-20L,
146 -1.8527916071632873716786E-20L,
147 1.8950325588932570796551E-20L,
148 1.5564775779538780478155E-20L,
149 6.0859793637556860974380E-21L,
150 -2.0208749253662532228949E-20L,
151 1.4966292219224761844552E-20L,
152 3.3540909728056476875639E-21L,
153 -8.6987564101742849540743E-22L,
154 -1.2327176863327626135542E-20L,
155 0.0000000000000000000000E0L
,
159 * on the interval -1/32 <= x <= 0
161 static const long double R
[] = {
162 1.5089970579127659901157E-5L,
163 1.5402715328927013076125E-4L,
164 1.3333556028915671091390E-3L,
165 9.6181291046036762031786E-3L,
166 5.5504108664798463044015E-2L,
167 2.4022650695910062854352E-1L,
168 6.9314718055994530931447E-1L,
171 #define douba(k) A[k]
172 #define doubb(k) B[k]
173 #define MEXP (NXT*16384.0L)
174 /* The following if denormal numbers are supported, else -MEXP: */
175 #define MNEXP (-NXT*(16384.0L+64.0L))
177 #define LOG2EA 0.44269504088896340735992L
189 static const long double MAXLOGL
= 1.1356523406294143949492E4L
;
190 static const long double MINLOGL
= -1.13994985314888605586758E4L
;
191 static const long double LOGE2L
= 6.9314718055994530941723E-1L;
193 static const long double huge
= 0x1p
10000L;
194 #if 0 /* XXX Prevent gcc from erroneously constant folding this. */
195 static const long double twom10000
= 0x1p
-10000L;
197 static const volatile long double twom10000
__attribute__ ((__section__(".rodata"))) = 0x1p
-10000L;
200 static long double reducl( long double );
201 static long double powil ( long double, int );
204 powl(long double x
, long double y
)
206 volatile long double z
;
207 long double w
, W
, Wa
, Wb
, ya
, yb
, u
;
208 /* double F, Fa, Fb, G, Ga, Gb, H, Ha, Hb */
209 int i
, nflg
, iyflg
, yoddint
;
226 if( !isfinite(y
) && x
== -1.0L )
233 if( x
> 0.0L && x
< 1.0L )
237 if( x
> -1.0L && x
< 0.0L )
244 if( x
> 0.0L && x
< 1.0L )
248 if( x
> -1.0L && x
< 0.0L )
259 /* Set iyflg to 1 if y is an integer. */
264 /* Test for odd integer y. */
269 ya
= floorl(0.5L * ya
);
270 yb
= 0.5L * fabsl(w
);
292 nflg
= 0; /* flag = 1 if x<0 raised to integer power */
299 if( signbit(x
) && yoddint
)
305 if( signbit(x
) && yoddint
)
310 return( 1.0L ); /* 0**0 */
312 return( 0.0L ); /* 0**y */
317 return (x
- x
) / (x
- x
); /* (x<0)**(non-int) is NaN */
322 /* Integer power of an integer. */
328 if( (w
== x
) && (fabsl(y
) < 32768.0) )
330 w
= powil( x
, (int) y
);
339 /* separate significand from exponent */
343 /* find significand in antilog table A[] */
347 if( x
<= douba(i
+8) )
349 if( x
<= douba(i
+4) )
351 if( x
<= douba(i
+2) )
358 /* Find (x - A[i])/A[i]
359 * in order to compute log(x/A[i]):
361 * log(x) = log( a x/a ) = log(a) + log(x/a)
363 * log(x/a) = log(1+v), v = x/a - 1 = (x-a)/a
370 /* rational approximation for log(1+v):
372 * log(1+v) = v - v**2/2 + v**3 P(v) / Q(v)
375 w
= x
* ( z
* __polevll( x
, (void *)P
, 3 ) / __p1evll( x
, (void *)Q
, 3 ) );
376 w
= w
- ldexpl( z
, -1 ); /* w - 0.5 * z */
378 /* Convert to base 2 logarithm:
379 * multiply by log2(e) = 1 + LOG2EA
386 /* Compute exponent term of the base 2 logarithm. */
388 w
= ldexpl( w
, -LNXT
); /* divide by NXT */
390 /* Now base 2 log of x is w + z. */
392 /* Multiply base 2 log by y, in extended precision. */
394 /* separate y into large part ya
395 * and small part yb less than 1/NXT
401 * = w*ya + w*yb + z*y
413 w
= ldexpl( Ga
+Ha
, LNXT
);
415 /* Test the power of 2 for overflow */
417 return (huge
* huge
); /* overflow */
420 return (twom10000
* twom10000
); /* underflow */
428 Hb
-= (1.0L/NXT
); /*0.0625L;*/
431 /* Now the product y * log2(x) = Hb + e/NXT.
433 * Compute base 2 exponential of Hb,
434 * where -0.0625 <= Hb <= 0.
436 z
= Hb
* __polevll( Hb
, (void *)R
, 6 ); /* z = 2**Hb - 1 */
438 /* Express e/NXT as an integer plus a negative number of (1/NXT)ths.
439 * Find lookup table entry for the fractional power of 2.
448 z
= w
* z
; /* 2**-e * ( 1 + (2**Hb-1) ) */
450 z
= ldexpl( z
, i
); /* multiply by integer power of 2 */
455 * find out if the integer exponent
462 z
= -z
; /* odd exponent */
469 /* Find a multiple of 1/NXT that is within 1/NXT of x. */
471 reducl(long double x
)
475 t
= ldexpl( x
, LNXT
);
477 t
= ldexpl( t
, -LNXT
);
483 * Real raised to integer power, long double precision
489 * long double x, y, powil();
498 * Returns argument x raised to the nth power.
499 * The routine efficiently decomposes n as a sum of powers of
500 * two. The desired power is a product of two-to-the-kth
501 * powers of x. Thus to compute the 32767 power of x requires
502 * 28 multiplications instead of 32767 multiplications.
510 * arithmetic x domain n domain # trials peak rms
511 * IEEE .001,1000 -1022,1023 50000 4.3e-17 7.8e-18
512 * IEEE 1,2 -1022,1023 20000 3.9e-17 7.6e-18
513 * IEEE .99,1.01 0,8700 10000 3.6e-16 7.2e-17
515 * Returns MAXNUM on overflow, zero on underflow.
520 powil(long double x
, int nn
)
524 int n
, e
, sign
, asign
, lx
;
560 /* Overflow detection */
562 /* Calculate approximate logarithm of answer */
564 s
= frexpl( s
, &lx
);
566 if( (e
== 0) || (e
> 64) || (e
< -64) )
568 s
= (s
- 7.0710678118654752e-1L) / (s
+ 7.0710678118654752e-1L);
569 s
= (2.9142135623730950L * s
- 0.5L + lx
) * nn
* LOGE2L
;
577 return (huge
* huge
); /* overflow */
580 return (twom10000
* twom10000
); /* underflow */
581 /* Handle tiny denormal answer, but with less accuracy
582 * since roundoff error in 1.0/x will be amplified.
583 * The precise demarcation should be the gradual underflow threshold.
585 if( s
< (-MAXLOGL
+2.0L) )
591 /* First bit of the power */
605 ww
= ww
* ww
; /* arg to the 2-to-the-kth power */
606 if( n
& 1 ) /* if that bit is set, then include in product */
612 y
= -y
; /* odd power of negative number */