]>
Commit | Line | Data |
---|---|---|
652df710 | 1 | /* Quad-precision floating point e^x. |
b168057a | 2 | Copyright (C) 1999-2015 Free Software Foundation, Inc. |
652df710 UD |
3 | This file is part of the GNU C Library. |
4 | Contributed by Jakub Jelinek <jj@ultra.linux.cz> | |
5 | Partly based on double-precision code | |
6 | by Geoffrey Keating <geoffk@ozemail.com.au> | |
7 | ||
8 | The GNU C Library is free software; you can redistribute it and/or | |
41bdb6e2 AJ |
9 | modify it under the terms of the GNU Lesser General Public |
10 | License as published by the Free Software Foundation; either | |
11 | version 2.1 of the License, or (at your option) any later version. | |
652df710 UD |
12 | |
13 | The GNU C Library is distributed in the hope that it will be useful, | |
14 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
41bdb6e2 | 16 | Lesser General Public License for more details. |
652df710 | 17 | |
41bdb6e2 | 18 | You should have received a copy of the GNU Lesser General Public |
59ba27a6 PE |
19 | License along with the GNU C Library; if not, see |
20 | <http://www.gnu.org/licenses/>. */ | |
652df710 UD |
21 | |
22 | /* The basic design here is from | |
23 | Abraham Ziv, "Fast Evaluation of Elementary Mathematical Functions with | |
24 | Correctly Rounded Last Bit", ACM Trans. Math. Soft., 17 (3), September 1991, | |
25 | pp. 410-423. | |
26 | ||
27 | We work with number pairs where the first number is the high part and | |
28 | the second one is the low part. Arithmetic with the high part numbers must | |
29 | be exact, without any roundoff errors. | |
30 | ||
31 | The input value, X, is written as | |
32 | X = n * ln(2)_0 + arg1[t1]_0 + arg2[t2]_0 + x | |
33 | - n * ln(2)_1 + arg1[t1]_1 + arg2[t2]_1 + xl | |
34 | ||
35 | where: | |
36 | - n is an integer, 16384 >= n >= -16495; | |
37 | - ln(2)_0 is the first 93 bits of ln(2), and |ln(2)_0-ln(2)-ln(2)_1| < 2^-205 | |
38 | - t1 is an integer, 89 >= t1 >= -89 | |
39 | - t2 is an integer, 65 >= t2 >= -65 | |
40 | - |arg1[t1]-t1/256.0| < 2^-53 | |
41 | - |arg2[t2]-t2/32768.0| < 2^-53 | |
42 | - x + xl is whatever is left, |x + xl| < 2^-16 + 2^-53 | |
43 | ||
44 | Then e^x is approximated as | |
45 | ||
46 | e^x = 2^n_1 ( 2^n_0 e^(arg1[t1]_0 + arg1[t1]_1) e^(arg2[t2]_0 + arg2[t2]_1) | |
47 | + 2^n_0 e^(arg1[t1]_0 + arg1[t1]_1) e^(arg2[t2]_0 + arg2[t2]_1) | |
48 | * p (x + xl + n * ln(2)_1)) | |
49 | where: | |
50 | - p(x) is a polynomial approximating e(x)-1 | |
51 | - e^(arg1[t1]_0 + arg1[t1]_1) is obtained from a table | |
52 | - e^(arg2[t2]_0 + arg2[t2]_1) likewise | |
53 | - n_1 + n_0 = n, so that |n_0| < -LDBL_MIN_EXP-1. | |
54 | ||
55 | If it happens that n_1 == 0 (this is the usual case), that multiplication | |
56 | is omitted. | |
57 | */ | |
58 | ||
59 | #ifndef _GNU_SOURCE | |
60 | #define _GNU_SOURCE | |
61 | #endif | |
62 | #include <float.h> | |
63 | #include <ieee754.h> | |
64 | #include <math.h> | |
65 | #include <fenv.h> | |
66 | #include <inttypes.h> | |
67 | #include <math_private.h> | |
e853ea00 | 68 | #include <stdlib.h> |
652df710 UD |
69 | #include "t_expl.h" |
70 | ||
71 | static const long double C[] = { | |
72 | /* Smallest integer x for which e^x overflows. */ | |
73 | #define himark C[0] | |
74 | 11356.523406294143949491931077970765L, | |
bcf01e6d | 75 | |
652df710 UD |
76 | /* Largest integer x for which e^x underflows. */ |
77 | #define lomark C[1] | |
78 | -11433.4627433362978788372438434526231L, | |
79 | ||
80 | /* 3x2^96 */ | |
81 | #define THREEp96 C[2] | |
82 | 59421121885698253195157962752.0L, | |
83 | ||
84 | /* 3x2^103 */ | |
85 | #define THREEp103 C[3] | |
86 | 30423614405477505635920876929024.0L, | |
87 | ||
88 | /* 3x2^111 */ | |
89 | #define THREEp111 C[4] | |
90 | 7788445287802241442795744493830144.0L, | |
91 | ||
92 | /* 1/ln(2) */ | |
93 | #define M_1_LN2 C[5] | |
94 | 1.44269504088896340735992468100189204L, | |
95 | ||
96 | /* first 93 bits of ln(2) */ | |
97 | #define M_LN2_0 C[6] | |
98 | 0.693147180559945309417232121457981864L, | |
99 | ||
100 | /* ln2_0 - ln(2) */ | |
101 | #define M_LN2_1 C[7] | |
102 | -1.94704509238074995158795957333327386E-31L, | |
103 | ||
104 | /* very small number */ | |
105 | #define TINY C[8] | |
106 | 1.0e-4900L, | |
107 | ||
108 | /* 2^16383 */ | |
109 | #define TWO16383 C[9] | |
110 | 5.94865747678615882542879663314003565E+4931L, | |
111 | ||
112 | /* 256 */ | |
113 | #define TWO8 C[10] | |
114 | 256.0L, | |
115 | ||
116 | /* 32768 */ | |
117 | #define TWO15 C[11] | |
118 | 32768.0L, | |
119 | ||
382466e0 | 120 | /* Chebyshev polynom coefficients for (exp(x)-1)/x */ |
652df710 UD |
121 | #define P1 C[12] |
122 | #define P2 C[13] | |
123 | #define P3 C[14] | |
124 | #define P4 C[15] | |
125 | #define P5 C[16] | |
126 | #define P6 C[17] | |
127 | 0.5L, | |
128 | 1.66666666666666666666666666666666683E-01L, | |
129 | 4.16666666666666666666654902320001674E-02L, | |
130 | 8.33333333333333333333314659767198461E-03L, | |
131 | 1.38888888889899438565058018857254025E-03L, | |
132 | 1.98412698413981650382436541785404286E-04L, | |
133 | }; | |
134 | ||
135 | long double | |
136 | __ieee754_expl (long double x) | |
137 | { | |
138 | /* Check for usual case. */ | |
139 | if (isless (x, himark) && isgreater (x, lomark)) | |
140 | { | |
141 | int tval1, tval2, unsafe, n_i; | |
142 | long double x22, n, t, result, xl; | |
143 | union ieee854_long_double ex2_u, scale_u; | |
144 | fenv_t oldenv; | |
145 | ||
146 | feholdexcept (&oldenv); | |
147 | #ifdef FE_TONEAREST | |
148 | fesetround (FE_TONEAREST); | |
149 | #endif | |
150 | ||
151 | /* Calculate n. */ | |
152 | n = x * M_1_LN2 + THREEp111; | |
153 | n -= THREEp111; | |
154 | x = x - n * M_LN2_0; | |
155 | xl = n * M_LN2_1; | |
156 | ||
157 | /* Calculate t/256. */ | |
158 | t = x + THREEp103; | |
159 | t -= THREEp103; | |
160 | ||
161 | /* Compute tval1 = t. */ | |
162 | tval1 = (int) (t * TWO8); | |
163 | ||
164 | x -= __expl_table[T_EXPL_ARG1+2*tval1]; | |
165 | xl -= __expl_table[T_EXPL_ARG1+2*tval1+1]; | |
166 | ||
167 | /* Calculate t/32768. */ | |
168 | t = x + THREEp96; | |
169 | t -= THREEp96; | |
170 | ||
171 | /* Compute tval2 = t. */ | |
172 | tval2 = (int) (t * TWO15); | |
173 | ||
174 | x -= __expl_table[T_EXPL_ARG2+2*tval2]; | |
175 | xl -= __expl_table[T_EXPL_ARG2+2*tval2+1]; | |
176 | ||
177 | x = x + xl; | |
178 | ||
179 | /* Compute ex2 = 2^n_0 e^(argtable[tval1]) e^(argtable[tval2]). */ | |
180 | ex2_u.d = __expl_table[T_EXPL_RES1 + tval1] | |
181 | * __expl_table[T_EXPL_RES2 + tval2]; | |
182 | n_i = (int)n; | |
183 | /* 'unsafe' is 1 iff n_1 != 0. */ | |
7e6424e3 | 184 | unsafe = abs(n_i) >= 15000; |
652df710 UD |
185 | ex2_u.ieee.exponent += n_i >> unsafe; |
186 | ||
187 | /* Compute scale = 2^n_1. */ | |
188 | scale_u.d = 1.0L; | |
189 | scale_u.ieee.exponent += n_i - (n_i >> unsafe); | |
190 | ||
191 | /* Approximate e^x2 - 1, using a seventh-degree polynomial, | |
192 | with maximum error in [-2^-16-2^-53,2^-16+2^-53] | |
193 | less than 4.8e-39. */ | |
194 | x22 = x + x*x*(P1+x*(P2+x*(P3+x*(P4+x*(P5+x*P6))))); | |
195 | ||
196 | /* Return result. */ | |
197 | fesetenv (&oldenv); | |
198 | ||
199 | result = x22 * ex2_u.d + ex2_u.d; | |
200 | ||
201 | /* Now we can test whether the result is ultimate or if we are unsure. | |
202 | In the later case we should probably call a mpn based routine to give | |
203 | the ultimate result. | |
204 | Empirically, this routine is already ultimate in about 99.9986% of | |
205 | cases, the test below for the round to nearest case will be false | |
206 | in ~ 99.9963% of cases. | |
207 | Without proc2 routine maximum error which has been seen is | |
208 | 0.5000262 ulp. | |
209 | ||
210 | union ieee854_long_double ex3_u; | |
211 | ||
212 | #ifdef FE_TONEAREST | |
213 | fesetround (FE_TONEAREST); | |
214 | #endif | |
215 | ex3_u.d = (result - ex2_u.d) - x22 * ex2_u.d; | |
216 | ex2_u.d = result; | |
217 | ex3_u.ieee.exponent += LDBL_MANT_DIG + 15 + IEEE854_LONG_DOUBLE_BIAS | |
350635a5 | 218 | - ex2_u.ieee.exponent; |
652df710 UD |
219 | n_i = abs (ex3_u.d); |
220 | n_i = (n_i + 1) / 2; | |
221 | fesetenv (&oldenv); | |
222 | #ifdef FE_TONEAREST | |
223 | if (fegetround () == FE_TONEAREST) | |
224 | n_i -= 0x4000; | |
225 | #endif | |
226 | if (!n_i) { | |
227 | return __ieee754_expl_proc2 (origx); | |
228 | } | |
229 | */ | |
230 | if (!unsafe) | |
231 | return result; | |
232 | else | |
233 | return result * scale_u.d; | |
234 | } | |
235 | /* Exceptional cases: */ | |
236 | else if (isless (x, himark)) | |
237 | { | |
d81f90cc | 238 | if (isinf (x)) |
652df710 UD |
239 | /* e^-inf == 0, with no error. */ |
240 | return 0; | |
241 | else | |
242 | /* Underflow */ | |
243 | return TINY * TINY; | |
244 | } | |
245 | else | |
246 | /* Return x, if x is a NaN or Inf; or overflow, otherwise. */ | |
247 | return TWO16383*x; | |
248 | } | |
bcf01e6d | 249 | strong_alias (__ieee754_expl, __expl_finite) |