2 * Copyright (c) 2008 Stephen L. Moshier <steve@moshier.net>
4 * Permission to use, copy, modify, and distribute this software for any
5 * purpose with or without fee is hereby granted, provided that the above
6 * copyright notice and this permission notice appear in all copies.
8 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
9 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
10 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
11 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
12 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
13 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
14 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
19 #include "math_private.h"
22 * Polynomial evaluator:
23 * P[0] x^n + P[1] x^(n-1) + ... + P[n]
25 static inline long double
26 __polevll(long double x, long double *PP, int n)
41 * Polynomial evaluator:
42 * x^n + P[0] x^(n-1) + P[1] x^(n-2) + ... + P[n]
44 static inline long double
45 __p1evll(long double x, long double *PP, int n)
62 * Power function, long double precision
68 * long double x, y, z, powl();
76 * Computes x raised to the yth power. Analytically,
78 * x**y = exp( y log(x) ).
80 * Following Cody and Waite, this program uses a lookup table
81 * of 2**-i/32 and pseudo extended precision arithmetic to
82 * obtain several extra bits of accuracy in both the logarithm
83 * and the exponential.
89 * The relative error of pow(x,y) can be estimated
90 * by y dl ln(2), where dl is the absolute error of
91 * the internally computed base 2 logarithm. At the ends
92 * of the approximation interval the logarithm equal 1/32
93 * and its relative error is about 1 lsb = 1.1e-19. Hence
94 * the predicted relative error in the result is 2.3e-21 y .
97 * arithmetic domain # trials peak rms
99 * IEEE +-1000 40000 2.8e-18 3.7e-19
100 * .001 < x < 1000, with log(x) uniformly distributed.
101 * -1000 < y < 1000, y uniformly distributed.
103 * IEEE 0,8700 60000 6.5e-18 1.0e-18
104 * 0.99 < x < 1.01, 0 < y < 8700, uniformly distributed.
109 * message condition value returned
110 * pow overflow x**y > MAXNUM INFINITY
111 * pow underflow x**y < 1/MAXNUM 0.0
112 * pow domain x<0 and y noninteger 0.0
119 #include "math_private.h"
123 /* log2(Table size) */
126 /* log(1+x) = x - .5x^2 + x^3 * P(z)/Q(z)
127 * on the domain 2^(-1/32) - 1 <= x <= 2^(1/32) - 1
129 static long double P[] = {
130 8.3319510773868690346226E-4L,
131 4.9000050881978028599627E-1L,
132 1.7500123722550302671919E0L,
133 1.4000100839971580279335E0L,
135 static long double Q[] = {
136 /* 1.0000000000000000000000E0L,*/
137 5.2500282295834889175431E0L,
138 8.4000598057587009834666E0L,
139 4.2000302519914740834728E0L,
141 /* A[i] = 2^(-i/32), rounded to IEEE long double precision.
142 * If i is even, A[i] + B[i/2] gives additional accuracy.
144 static long double A[33] = {
145 1.0000000000000000000000E0L,
146 9.7857206208770013448287E-1L,
147 9.5760328069857364691013E-1L,
148 9.3708381705514995065011E-1L,
149 9.1700404320467123175367E-1L,
150 8.9735453750155359320742E-1L,
151 8.7812608018664974155474E-1L,
152 8.5930964906123895780165E-1L,
153 8.4089641525371454301892E-1L,
154 8.2287773907698242225554E-1L,
155 8.0524516597462715409607E-1L,
156 7.8799042255394324325455E-1L,
157 7.7110541270397041179298E-1L,
158 7.5458221379671136985669E-1L,
159 7.3841307296974965571198E-1L,
160 7.2259040348852331001267E-1L,
161 7.0710678118654752438189E-1L,
162 6.9195494098191597746178E-1L,
163 6.7712777346844636413344E-1L,
164 6.6261832157987064729696E-1L,
165 6.4841977732550483296079E-1L,
166 6.3452547859586661129850E-1L,
167 6.2092890603674202431705E-1L,
168 6.0762367999023443907803E-1L,
169 5.9460355750136053334378E-1L,
170 5.8186242938878875689693E-1L,
171 5.6939431737834582684856E-1L,
172 5.5719337129794626814472E-1L,
173 5.4525386633262882960438E-1L,
174 5.3357020033841180906486E-1L,
175 5.2213689121370692017331E-1L,
176 5.1094857432705833910408E-1L,
177 5.0000000000000000000000E-1L,
179 static long double B[17] = {
180 0.0000000000000000000000E0L,
181 2.6176170809902549338711E-20L,
182 -1.0126791927256478897086E-20L,
183 1.3438228172316276937655E-21L,
184 1.2207982955417546912101E-20L,
185 -6.3084814358060867200133E-21L,
186 1.3164426894366316434230E-20L,
187 -1.8527916071632873716786E-20L,
188 1.8950325588932570796551E-20L,
189 1.5564775779538780478155E-20L,
190 6.0859793637556860974380E-21L,
191 -2.0208749253662532228949E-20L,
192 1.4966292219224761844552E-20L,
193 3.3540909728056476875639E-21L,
194 -8.6987564101742849540743E-22L,
195 -1.2327176863327626135542E-20L,
196 0.0000000000000000000000E0L,
200 * on the interval -1/32 <= x <= 0
202 static long double R[] = {
203 1.5089970579127659901157E-5L,
204 1.5402715328927013076125E-4L,
205 1.3333556028915671091390E-3L,
206 9.6181291046036762031786E-3L,
207 5.5504108664798463044015E-2L,
208 2.4022650695910062854352E-1L,
209 6.9314718055994530931447E-1L,
212 #define douba(k) A[k]
213 #define doubb(k) B[k]
214 #define MEXP (NXT*16384.0L)
215 /* The following if denormal numbers are supported, else -MEXP: */
216 #define MNEXP (-NXT*(16384.0L+64.0L))
218 #define LOG2EA 0.44269504088896340735992L
230 static const long double MAXLOGL = 1.1356523406294143949492E4L;
231 static const long double MINLOGL = -1.13994985314888605586758E4L;
232 static const long double LOGE2L = 6.9314718055994530941723E-1L;
233 static volatile long double z;
234 static long double w, W, Wa, Wb, ya, yb, u;
235 static const long double huge = 0x1p10000L;
236 #if 0 /* XXX Prevent gcc from erroneously constant folding this. */
237 static const long double twom10000 = 0x1p-10000L;
239 static volatile long double twom10000 = 0x1p-10000L;
242 static long double reducl( long double );
243 static long double powil ( long double, int );
246 powl(long double x, long double y)
248 /* double F, Fa, Fb, G, Ga, Gb, H, Ha, Hb */
249 int i, nflg, iyflg, yoddint;
259 return ( nan_mix(x, y) );
261 return ( nan_mix(x, y) );
266 if( !isfinite(y) && x == -1.0L )
273 if( x > 0.0L && x < 1.0L )
277 if( x > -1.0L && x < 0.0L )
284 if( x > 0.0L && x < 1.0L )
288 if( x > -1.0L && x < 0.0L )
299 /* Set iyflg to 1 if y is an integer. */
304 /* Test for odd integer y. */
309 ya = floorl(0.5L * ya);
310 yb = 0.5L * fabsl(w);
332 nflg = 0; /* flag = 1 if x<0 raised to integer power */
339 if( signbit(x) && yoddint )
345 if( signbit(x) && yoddint )
350 return( 1.0L ); /* 0**0 */
352 return( 0.0L ); /* 0**y */
357 return (x - x) / (x - x); /* (x<0)**(non-int) is NaN */
362 /* Integer power of an integer. */
368 if( (w == x) && (fabsl(y) < 32768.0) )
370 w = powil( x, (int) y );
379 /* separate significand from exponent */
383 /* find significand in antilog table A[] */
387 if( x <= douba(i+8) )
389 if( x <= douba(i+4) )
391 if( x <= douba(i+2) )
398 /* Find (x - A[i])/A[i]
399 * in order to compute log(x/A[i]):
401 * log(x) = log( a x/a ) = log(a) + log(x/a)
403 * log(x/a) = log(1+v), v = x/a - 1 = (x-a)/a
410 /* rational approximation for log(1+v):
412 * log(1+v) = v - v**2/2 + v**3 P(v) / Q(v)
415 w = x * ( z * __polevll( x, P, 3 ) / __p1evll( x, Q, 3 ) );
416 w = w - ldexpl( z, -1 ); /* w - 0.5 * z */
418 /* Convert to base 2 logarithm:
419 * multiply by log2(e) = 1 + LOG2EA
426 /* Compute exponent term of the base 2 logarithm. */
428 w = ldexpl( w, -LNXT ); /* divide by NXT */
430 /* Now base 2 log of x is w + z. */
432 /* Multiply base 2 log by y, in extended precision. */
434 /* separate y into large part ya
435 * and small part yb less than 1/NXT
441 * = w*ya + w*yb + z*y
453 w = ldexpl( Ga+Ha, LNXT );
455 /* Test the power of 2 for overflow */
457 return (huge * huge); /* overflow */
460 return (twom10000 * twom10000); /* underflow */
468 Hb -= (1.0L/NXT); /*0.0625L;*/
471 /* Now the product y * log2(x) = Hb + e/NXT.
473 * Compute base 2 exponential of Hb,
474 * where -0.0625 <= Hb <= 0.
476 z = Hb * __polevll( Hb, R, 6 ); /* z = 2**Hb - 1 */
478 /* Express e/NXT as an integer plus a negative number of (1/NXT)ths.
479 * Find lookup table entry for the fractional power of 2.
488 z = w * z; /* 2**-e * ( 1 + (2**Hb-1) ) */
490 z = ldexpl( z, i ); /* multiply by integer power of 2 */
495 * find out if the integer exponent
502 z = -z; /* odd exponent */
509 /* Find a multiple of 1/NXT that is within 1/NXT of x. */
510 static inline long double
511 reducl(long double x)
515 t = ldexpl( x, LNXT );
517 t = ldexpl( t, -LNXT );
523 * Real raised to integer power, long double precision
529 * long double x, y, powil();
538 * Returns argument x raised to the nth power.
539 * The routine efficiently decomposes n as a sum of powers of
540 * two. The desired power is a product of two-to-the-kth
541 * powers of x. Thus to compute the 32767 power of x requires
542 * 28 multiplications instead of 32767 multiplications.
550 * arithmetic x domain n domain # trials peak rms
551 * IEEE .001,1000 -1022,1023 50000 4.3e-17 7.8e-18
552 * IEEE 1,2 -1022,1023 20000 3.9e-17 7.6e-18
553 * IEEE .99,1.01 0,8700 10000 3.6e-16 7.2e-17
555 * Returns MAXNUM on overflow, zero on underflow.
560 powil(long double x, int nn)
564 int n, e, sign, asign, lx;
600 /* Overflow detection */
602 /* Calculate approximate logarithm of answer */
604 s = frexpl( s, &lx );
606 if( (e == 0) || (e > 64) || (e < -64) )
608 s = (s - 7.0710678118654752e-1L) / (s + 7.0710678118654752e-1L);
609 s = (2.9142135623730950L * s - 0.5L + lx) * nn * LOGE2L;
617 return (huge * huge); /* overflow */
620 return (twom10000 * twom10000); /* underflow */
621 /* Handle tiny denormal answer, but with less accuracy
622 * since roundoff error in 1.0/x will be amplified.
623 * The precise demarcation should be the gradual underflow threshold.
625 if( s < (-MAXLOGL+2.0L) )
631 /* First bit of the power */
645 ww = ww * ww; /* arg to the 2-to-the-kth power */
646 if( n & 1 ) /* if that bit is set, then include in product */
652 y = -y; /* odd power of negative number */