2001-03-12 00:04:52 +00:00
|
|
|
/*
|
|
|
|
* IBM Accurate Mathematical Library
|
2002-07-06 06:36:39 +00:00
|
|
|
* written by International Business Machines Corp.
|
2018-01-01 00:32:25 +00:00
|
|
|
* Copyright (C) 2001-2018 Free Software Foundation, Inc.
|
2001-03-12 00:04:52 +00:00
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or modify
|
|
|
|
* it under the terms of the GNU Lesser General Public License as published by
|
2002-08-26 22:40:48 +00:00
|
|
|
* the Free Software Foundation; either version 2.1 of the License, or
|
2001-03-12 00:04:52 +00:00
|
|
|
* (at your option) any later version.
|
|
|
|
*
|
|
|
|
* This program is distributed in the hope that it will be useful,
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
2002-08-20 21:51:55 +00:00
|
|
|
* GNU Lesser General Public License for more details.
|
2001-03-12 00:04:52 +00:00
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU Lesser General Public License
|
2012-02-09 23:18:22 +00:00
|
|
|
* along with this program; if not, see <http://www.gnu.org/licenses/>.
|
2001-03-12 00:04:52 +00:00
|
|
|
*/
|
|
|
|
/***************************************************************************/
|
|
|
|
/* MODULE_NAME:uexp.c */
|
|
|
|
/* */
|
|
|
|
/* FUNCTION:uexp */
|
|
|
|
/* exp1 */
|
|
|
|
/* */
|
|
|
|
/* FILES NEEDED:dla.h endian.h mpa.h mydefs.h uexp.h */
|
2017-12-19 18:11:37 +00:00
|
|
|
/* mpa.c mpexp.x slowexp.c */
|
2001-03-12 00:04:52 +00:00
|
|
|
/* */
|
|
|
|
/* An ultimate exp routine. Given an IEEE double machine number x */
|
|
|
|
/* it computes the correctly rounded (to nearest) value of e^x */
|
|
|
|
/* Assumption: Machine arithmetic operations are performed in */
|
|
|
|
/* round to nearest mode of IEEE 754 standard. */
|
|
|
|
/* */
|
|
|
|
/***************************************************************************/
|
|
|
|
|
2014-03-24 22:00:32 +00:00
|
|
|
#include <math.h>
|
2001-03-12 00:04:52 +00:00
|
|
|
#include "endian.h"
|
2001-05-12 14:32:12 +00:00
|
|
|
#include "uexp.h"
|
2001-03-12 00:04:52 +00:00
|
|
|
#include "mydefs.h"
|
|
|
|
#include "MathLib.h"
|
2017-12-19 18:11:37 +00:00
|
|
|
#include "uexp.tbl"
|
|
|
|
#include <math_private.h>
|
2012-03-02 15:12:53 +00:00
|
|
|
#include <fenv.h>
|
2013-12-03 21:49:56 +00:00
|
|
|
#include <float.h>
|
2001-05-12 14:32:12 +00:00
|
|
|
|
2017-12-19 18:11:37 +00:00
|
|
|
#ifndef SECTION
|
|
|
|
# define SECTION
|
|
|
|
#endif
|
2011-10-25 04:56:33 +00:00
|
|
|
|
2017-12-19 18:11:37 +00:00
|
|
|
double __slowexp (double);
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2017-12-19 18:11:37 +00:00
|
|
|
/* An ultimate exp routine. Given an IEEE double machine number x it computes
|
|
|
|
the correctly rounded (to nearest) value of e^x. */
|
2011-10-25 04:56:33 +00:00
|
|
|
double
|
2017-12-19 18:11:37 +00:00
|
|
|
SECTION
|
|
|
|
__ieee754_exp (double x)
|
2013-10-08 10:52:28 +00:00
|
|
|
{
|
2017-12-19 18:11:37 +00:00
|
|
|
double bexp, t, eps, del, base, y, al, bet, res, rem, cor;
|
|
|
|
mynumber junk1, junk2, binexp = {{0, 0}};
|
|
|
|
int4 i, j, m, n, ex;
|
2012-03-02 15:12:53 +00:00
|
|
|
double retval;
|
Improve __ieee754_exp() performance by greater than 5x on sparc/x86.
These changes will be active for all platforms that don't provide
their own exp() routines. They will also be active for ieee754
versions of ccos, ccosh, cosh, csin, csinh, sinh, exp10, gamma, and
erf.
Typical performance gains is typically around 5x when measured on
Sparc s7 for common values between exp(1) and exp(40).
Using the glibc perf tests on sparc,
sparc (nsec) x86 (nsec)
old new old new
max 17629 395 5173 144
min 399 54 15 13
mean 5317 200 1349 23
The extreme max times for the old (ieee754) exp are due to the
multiprecision computation in the old algorithm when the true value is
very near 0.5 ulp away from an value representable in double
precision. The new algorithm does not take special measures for those
cases. The current glibc exp perf tests overrepresent those values.
Informal testing suggests approximately one in 200 cases might
invoke the high cost computation. The performance advantage of the new
algorithm for other values is still large but not as large as indicated
by the chart above.
Glibc correctness tests for exp() and expf() were run. Within the
test suite 3 input values were found to cause 1 bit differences (ulp)
when "FE_TONEAREST" rounding mode is set. No differences in exp() were
seen for the tested values for the other rounding modes.
Typical example:
exp(-0x1.760cd2p+0) (-1.46113312244415283203125)
new code: 2.31973271630014299393707e-01 0x1.db14cd799387ap-3
old code: 2.31973271630014271638132e-01 0x1.db14cd7993879p-3
exp = 2.31973271630014285508337 (high precision)
Old delta: off by 0.49 ulp
New delta: off by 0.51 ulp
In addition, because ieee754_exp() is used by other routines, cexp()
showed test results with very small imaginary input values where the
imaginary portion of the result was off by 3 ulp when in upward
rounding mode, but not in the other rounding modes. For x86, tgamma
showed a few values where the ulp increased to 6 (max ulp for tgamma
is 5). Sparc tgamma did not show these failures. I presume the tgamma
differences are due to compiler optimization differences within the
gamma function.The gamma function is known to be difficult to compute
accurately.
* sysdeps/ieee754/dbl-64/e_exp.c: Include <math-svid-compat.h> and
<errno.h>. Include "eexp.tbl".
(half): New constant.
(one): Likewise.
(__ieee754_exp): Rewrite.
(__slowexp): Remove prototype.
* sysdeps/ieee754/dbl-64/eexp.tbl: New file.
* sysdeps/ieee754/dbl-64/slowexp.c: Remove file.
* sysdeps/i386/fpu/slowexp.c: Likewise.
* sysdeps/ia64/fpu/slowexp.c: Likewise.
* sysdeps/m68k/m680x0/fpu/slowexp.c: Likewise.
* sysdeps/x86_64/fpu/multiarch/slowexp-avx.c: Likewise.
* sysdeps/x86_64/fpu/multiarch/slowexp-fma.c: Likewise.
* sysdeps/x86_64/fpu/multiarch/slowexp-fma4.c: Likewise.
* sysdeps/generic/math_private.h (__slowexp): Remove prototype.
* sysdeps/ieee754/dbl-64/e_pow.c: Remove mention of slowexp.c in
comment.
* sysdeps/powerpc/power4/fpu/Makefile [$(subdir) = math]
(CPPFLAGS-slowexp.c): Remove variable.
* sysdeps/x86_64/fpu/multiarch/Makefile (libm-sysdep_routines):
Remove slowexp-fma, slowexp-fma4 and slowexp-avx.
(CFLAGS-slowexp-fma.c): Remove variable.
(CFLAGS-slowexp-fma4.c): Likewise.
(CFLAGS-slowexp-avx.c): Likewise.
* sysdeps/x86_64/fpu/multiarch/e_exp-avx.c (__slowexp): Do not
define as macro.
* sysdeps/x86_64/fpu/multiarch/e_exp-fma.c (__slowexp): Likewise.
* sysdeps/x86_64/fpu/multiarch/e_exp-fma4.c (__slowexp): Likewise.
* math/Makefile (type-double-routines): Remove slowexp.
* manual/probes.texi (slowexp_p6): Remove.
(slowexp_p32): Likewise.
2017-12-19 17:25:14 +00:00
|
|
|
|
2017-12-19 18:11:37 +00:00
|
|
|
{
|
|
|
|
SET_RESTORE_ROUND (FE_TONEAREST);
|
|
|
|
|
|
|
|
junk1.x = x;
|
|
|
|
m = junk1.i[HIGH_HALF];
|
|
|
|
n = m & hugeint;
|
|
|
|
|
|
|
|
if (n > smallint && n < bigint)
|
|
|
|
{
|
|
|
|
y = x * log2e.x + three51.x;
|
|
|
|
bexp = y - three51.x; /* multiply the result by 2**bexp */
|
|
|
|
|
|
|
|
junk1.x = y;
|
|
|
|
|
|
|
|
eps = bexp * ln_two2.x; /* x = bexp*ln(2) + t - eps */
|
|
|
|
t = x - bexp * ln_two1.x;
|
|
|
|
|
|
|
|
y = t + three33.x;
|
|
|
|
base = y - three33.x; /* t rounded to a multiple of 2**-18 */
|
|
|
|
junk2.x = y;
|
|
|
|
del = (t - base) - eps; /* x = bexp*ln(2) + base + del */
|
|
|
|
eps = del + del * del * (p3.x * del + p2.x);
|
|
|
|
|
|
|
|
binexp.i[HIGH_HALF] = (junk1.i[LOW_HALF] + 1023) << 20;
|
|
|
|
|
|
|
|
i = ((junk2.i[LOW_HALF] >> 8) & 0xfffffffe) + 356;
|
|
|
|
j = (junk2.i[LOW_HALF] & 511) << 1;
|
|
|
|
|
|
|
|
al = coar.x[i] * fine.x[j];
|
|
|
|
bet = ((coar.x[i] * fine.x[j + 1] + coar.x[i + 1] * fine.x[j])
|
|
|
|
+ coar.x[i + 1] * fine.x[j + 1]);
|
|
|
|
|
|
|
|
rem = (bet + bet * eps) + al * eps;
|
|
|
|
res = al + rem;
|
|
|
|
cor = (al - res) + rem;
|
|
|
|
if (res == (res + cor * err_0))
|
|
|
|
{
|
|
|
|
retval = res * binexp.x;
|
|
|
|
goto ret;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
retval = __slowexp (x);
|
|
|
|
goto ret;
|
|
|
|
} /*if error is over bound */
|
|
|
|
}
|
|
|
|
|
|
|
|
if (n <= smallint)
|
|
|
|
{
|
|
|
|
retval = 1.0;
|
|
|
|
goto ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (n >= badint)
|
|
|
|
{
|
|
|
|
if (n > infint)
|
|
|
|
{
|
|
|
|
retval = x + x;
|
|
|
|
goto ret;
|
|
|
|
} /* x is NaN */
|
|
|
|
if (n < infint)
|
|
|
|
{
|
|
|
|
if (x > 0)
|
|
|
|
goto ret_huge;
|
|
|
|
else
|
|
|
|
goto ret_tiny;
|
|
|
|
}
|
|
|
|
/* x is finite, cause either overflow or underflow */
|
|
|
|
if (junk1.i[LOW_HALF] != 0)
|
|
|
|
{
|
|
|
|
retval = x + x;
|
|
|
|
goto ret;
|
|
|
|
} /* x is NaN */
|
|
|
|
retval = (x > 0) ? inf.x : zero; /* |x| = inf; return either inf or 0 */
|
|
|
|
goto ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
y = x * log2e.x + three51.x;
|
|
|
|
bexp = y - three51.x;
|
|
|
|
junk1.x = y;
|
|
|
|
eps = bexp * ln_two2.x;
|
|
|
|
t = x - bexp * ln_two1.x;
|
|
|
|
y = t + three33.x;
|
|
|
|
base = y - three33.x;
|
|
|
|
junk2.x = y;
|
|
|
|
del = (t - base) - eps;
|
|
|
|
eps = del + del * del * (p3.x * del + p2.x);
|
|
|
|
i = ((junk2.i[LOW_HALF] >> 8) & 0xfffffffe) + 356;
|
|
|
|
j = (junk2.i[LOW_HALF] & 511) << 1;
|
|
|
|
al = coar.x[i] * fine.x[j];
|
|
|
|
bet = ((coar.x[i] * fine.x[j + 1] + coar.x[i + 1] * fine.x[j])
|
|
|
|
+ coar.x[i + 1] * fine.x[j + 1]);
|
|
|
|
rem = (bet + bet * eps) + al * eps;
|
|
|
|
res = al + rem;
|
|
|
|
cor = (al - res) + rem;
|
|
|
|
if (m >> 31)
|
|
|
|
{
|
|
|
|
ex = junk1.i[LOW_HALF];
|
|
|
|
if (res < 1.0)
|
|
|
|
{
|
|
|
|
res += res;
|
|
|
|
cor += cor;
|
|
|
|
ex -= 1;
|
|
|
|
}
|
|
|
|
if (ex >= -1022)
|
|
|
|
{
|
|
|
|
binexp.i[HIGH_HALF] = (1023 + ex) << 20;
|
|
|
|
if (res == (res + cor * err_0))
|
|
|
|
{
|
|
|
|
retval = res * binexp.x;
|
|
|
|
goto ret;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
retval = __slowexp (x);
|
|
|
|
goto check_uflow_ret;
|
|
|
|
} /*if error is over bound */
|
|
|
|
}
|
|
|
|
ex = -(1022 + ex);
|
|
|
|
binexp.i[HIGH_HALF] = (1023 - ex) << 20;
|
|
|
|
res *= binexp.x;
|
|
|
|
cor *= binexp.x;
|
|
|
|
eps = 1.0000000001 + err_0 * binexp.x;
|
|
|
|
t = 1.0 + res;
|
|
|
|
y = ((1.0 - t) + res) + cor;
|
|
|
|
res = t + y;
|
|
|
|
cor = (t - res) + y;
|
|
|
|
if (res == (res + eps * cor))
|
|
|
|
{
|
|
|
|
binexp.i[HIGH_HALF] = 0x00100000;
|
|
|
|
retval = (res - 1.0) * binexp.x;
|
|
|
|
goto check_uflow_ret;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
retval = __slowexp (x);
|
|
|
|
goto check_uflow_ret;
|
|
|
|
} /* if error is over bound */
|
|
|
|
check_uflow_ret:
|
|
|
|
if (retval < DBL_MIN)
|
|
|
|
{
|
|
|
|
double force_underflow = tiny * tiny;
|
|
|
|
math_force_eval (force_underflow);
|
|
|
|
}
|
|
|
|
if (retval == 0)
|
|
|
|
goto ret_tiny;
|
|
|
|
goto ret;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
binexp.i[HIGH_HALF] = (junk1.i[LOW_HALF] + 767) << 20;
|
|
|
|
if (res == (res + cor * err_0))
|
|
|
|
retval = res * binexp.x * t256.x;
|
|
|
|
else
|
|
|
|
retval = __slowexp (x);
|
|
|
|
if (isinf (retval))
|
|
|
|
goto ret_huge;
|
|
|
|
else
|
|
|
|
goto ret;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
ret:
|
|
|
|
return retval;
|
|
|
|
|
|
|
|
ret_huge:
|
|
|
|
return hhuge * hhuge;
|
|
|
|
|
|
|
|
ret_tiny:
|
|
|
|
return tiny * tiny;
|
2001-03-12 00:04:52 +00:00
|
|
|
}
|
2011-10-25 00:19:17 +00:00
|
|
|
#ifndef __ieee754_exp
|
2011-10-16 00:22:59 +00:00
|
|
|
strong_alias (__ieee754_exp, __exp_finite)
|
2011-10-25 00:19:17 +00:00
|
|
|
#endif
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
/* Compute e^(x+xx). The routine also receives bound of error of previous
|
|
|
|
calculation. If after computing exp the error exceeds the allowed bounds,
|
|
|
|
the routine returns a non-positive number. Otherwise it returns the
|
|
|
|
computed result, which is always positive. */
|
2011-10-25 04:56:33 +00:00
|
|
|
double
|
|
|
|
SECTION
|
2013-10-08 10:52:28 +00:00
|
|
|
__exp1 (double x, double xx, double error)
|
|
|
|
{
|
2001-03-12 00:04:52 +00:00
|
|
|
double bexp, t, eps, del, base, y, al, bet, res, rem, cor;
|
2013-10-08 10:52:28 +00:00
|
|
|
mynumber junk1, junk2, binexp = {{0, 0}};
|
|
|
|
int4 i, j, m, n, ex;
|
2001-03-12 00:04:52 +00:00
|
|
|
|
|
|
|
junk1.x = x;
|
|
|
|
m = junk1.i[HIGH_HALF];
|
2013-10-08 10:52:28 +00:00
|
|
|
n = m & hugeint; /* no sign */
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
if (n > smallint && n < bigint)
|
|
|
|
{
|
|
|
|
y = x * log2e.x + three51.x;
|
|
|
|
bexp = y - three51.x; /* multiply the result by 2**bexp */
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
junk1.x = y;
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
eps = bexp * ln_two2.x; /* x = bexp*ln(2) + t - eps */
|
|
|
|
t = x - bexp * ln_two1.x;
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
y = t + three33.x;
|
|
|
|
base = y - three33.x; /* t rounded to a multiple of 2**-18 */
|
|
|
|
junk2.x = y;
|
|
|
|
del = (t - base) + (xx - eps); /* x = bexp*ln(2) + base + del */
|
|
|
|
eps = del + del * del * (p3.x * del + p2.x);
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
binexp.i[HIGH_HALF] = (junk1.i[LOW_HALF] + 1023) << 20;
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
i = ((junk2.i[LOW_HALF] >> 8) & 0xfffffffe) + 356;
|
|
|
|
j = (junk2.i[LOW_HALF] & 511) << 1;
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
al = coar.x[i] * fine.x[j];
|
|
|
|
bet = ((coar.x[i] * fine.x[j + 1] + coar.x[i + 1] * fine.x[j])
|
|
|
|
+ coar.x[i + 1] * fine.x[j + 1]);
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
rem = (bet + bet * eps) + al * eps;
|
|
|
|
res = al + rem;
|
|
|
|
cor = (al - res) + rem;
|
|
|
|
if (res == (res + cor * (1.0 + error + err_1)))
|
|
|
|
return res * binexp.x;
|
|
|
|
else
|
|
|
|
return -10.0;
|
|
|
|
}
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
if (n <= smallint)
|
|
|
|
return 1.0; /* if x->0 e^x=1 */
|
|
|
|
|
|
|
|
if (n >= badint)
|
|
|
|
{
|
|
|
|
if (n > infint)
|
|
|
|
return (zero / zero); /* x is NaN, return invalid */
|
|
|
|
if (n < infint)
|
|
|
|
return ((x > 0) ? (hhuge * hhuge) : (tiny * tiny));
|
|
|
|
/* x is finite, cause either overflow or underflow */
|
|
|
|
if (junk1.i[LOW_HALF] != 0)
|
|
|
|
return (zero / zero); /* x is NaN */
|
|
|
|
return ((x > 0) ? inf.x : zero); /* |x| = inf; return either inf or 0 */
|
|
|
|
}
|
2001-03-12 00:04:52 +00:00
|
|
|
|
2013-10-08 10:52:28 +00:00
|
|
|
y = x * log2e.x + three51.x;
|
2001-03-12 00:04:52 +00:00
|
|
|
bexp = y - three51.x;
|
|
|
|
junk1.x = y;
|
2013-10-08 10:52:28 +00:00
|
|
|
eps = bexp * ln_two2.x;
|
|
|
|
t = x - bexp * ln_two1.x;
|
2001-03-12 00:04:52 +00:00
|
|
|
y = t + three33.x;
|
|
|
|
base = y - three33.x;
|
|
|
|
junk2.x = y;
|
2013-10-08 10:52:28 +00:00
|
|
|
del = (t - base) + (xx - eps);
|
|
|
|
eps = del + del * del * (p3.x * del + p2.x);
|
|
|
|
i = ((junk2.i[LOW_HALF] >> 8) & 0xfffffffe) + 356;
|
|
|
|
j = (junk2.i[LOW_HALF] & 511) << 1;
|
|
|
|
al = coar.x[i] * fine.x[j];
|
|
|
|
bet = ((coar.x[i] * fine.x[j + 1] + coar.x[i + 1] * fine.x[j])
|
|
|
|
+ coar.x[i + 1] * fine.x[j + 1]);
|
|
|
|
rem = (bet + bet * eps) + al * eps;
|
2001-03-12 00:04:52 +00:00
|
|
|
res = al + rem;
|
|
|
|
cor = (al - res) + rem;
|
2013-10-08 10:52:28 +00:00
|
|
|
if (m >> 31)
|
|
|
|
{
|
|
|
|
ex = junk1.i[LOW_HALF];
|
|
|
|
if (res < 1.0)
|
|
|
|
{
|
|
|
|
res += res;
|
|
|
|
cor += cor;
|
|
|
|
ex -= 1;
|
|
|
|
}
|
|
|
|
if (ex >= -1022)
|
|
|
|
{
|
|
|
|
binexp.i[HIGH_HALF] = (1023 + ex) << 20;
|
|
|
|
if (res == (res + cor * (1.0 + error + err_1)))
|
|
|
|
return res * binexp.x;
|
|
|
|
else
|
|
|
|
return -10.0;
|
|
|
|
}
|
|
|
|
ex = -(1022 + ex);
|
|
|
|
binexp.i[HIGH_HALF] = (1023 - ex) << 20;
|
|
|
|
res *= binexp.x;
|
|
|
|
cor *= binexp.x;
|
|
|
|
eps = 1.00000000001 + (error + err_1) * binexp.x;
|
|
|
|
t = 1.0 + res;
|
|
|
|
y = ((1.0 - t) + res) + cor;
|
|
|
|
res = t + y;
|
|
|
|
cor = (t - res) + y;
|
|
|
|
if (res == (res + eps * cor))
|
|
|
|
{
|
|
|
|
binexp.i[HIGH_HALF] = 0x00100000;
|
|
|
|
return (res - 1.0) * binexp.x;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
return -10.0;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
binexp.i[HIGH_HALF] = (junk1.i[LOW_HALF] + 767) << 20;
|
|
|
|
if (res == (res + cor * (1.0 + error + err_1)))
|
|
|
|
return res * binexp.x * t256.x;
|
|
|
|
else
|
|
|
|
return -10.0;
|
2001-03-12 00:04:52 +00:00
|
|
|
}
|
1996-03-05 21:41:30 +00:00
|
|
|
}
|