Merge powerpc slowpow.c into generic code
This commit is contained in:
Siddhesh Poyarekar
parent
e6ebd4a7d5
commit
4cc149fd8e
10
ChangeLog
10
ChangeLog
@ -1,5 +1,15 @@
|
||||
2013-03-07 Siddhesh Poyarekar <siddhesh@redhat.com>
|
||||
|
||||
* sysdeps/ieee754/dbl-64/slowpow.c [USE_LONG_DOUBLE_FOR_MP]
|
||||
(__slowpow): Use long double EXPL and LOGL functions to
|
||||
compute POW.
|
||||
* sysdeps/powerpc/powerpc32/power4/fpu/Makefile
|
||||
(CPPFLAGS-slowpow.c): Define USE_LONG_DOUBLE_FOR_MP.
|
||||
* sysdeps/powerpc/powerpc32/power4/fpu/slowpow.c: Remove.
|
||||
* sysdeps/powerpc/powerpc64/power4/fpu/Makefile
|
||||
(CPPFLAGS-slowpow.c): Define USE_LONG_DOUBLE_FOR_MP.
|
||||
* sysdeps/powerpc/powerpc64/power4/fpu/slowpow.c: Remove.
|
||||
|
||||
* sysdeps/powerpc/powerpc32/power4/fpu/mpa.c (__mul): Use
|
||||
intermediate variable to calculate exponent.
|
||||
(__sqr): Likewise.
|
||||
|
||||
@ -59,6 +59,23 @@ __slowpow (double x, double y, double z)
|
||||
if (res >= 0)
|
||||
return res;
|
||||
|
||||
/* Compute pow as long double. This is currently only used by powerpc, where
|
||||
one may get 106 bits of accuracy. */
|
||||
#ifdef USE_LONG_DOUBLE_FOR_MP
|
||||
long double ldw, ldz, ldpp;
|
||||
static const long double ldeps = 0x4.0p-96;
|
||||
|
||||
ldz = __ieee754_logl ((long double) x);
|
||||
ldw = (long double) y *ldz;
|
||||
ldpp = __ieee754_expl (ldw);
|
||||
res = (double) (ldpp + ldeps);
|
||||
res1 = (double) (ldpp - ldeps);
|
||||
|
||||
/* Return the result if it is accurate enough. */
|
||||
if (res == res1)
|
||||
return res;
|
||||
#endif
|
||||
|
||||
/* Or else, calculate using multiple precision. P = 10 implies accuracy of
|
||||
240 bits accuracy, since MP_NO has a radix of 2^24. */
|
||||
p = 10;
|
||||
|
||||
@ -2,4 +2,5 @@
|
||||
|
||||
ifeq ($(subdir),math)
|
||||
CFLAGS-mpa.c += --param max-unroll-times=4 -funroll-loops -fpeel-loops
|
||||
CPPFLAGS-slowpow.c += -DUSE_LONG_DOUBLE_FOR_MP=1
|
||||
endif
|
||||
|
||||
@ -1,93 +0,0 @@
|
||||
/*
|
||||
* IBM Accurate Mathematical Library
|
||||
* written by International Business Machines Corp.
|
||||
* Copyright (C) 2001-2013 Free Software Foundation, Inc.
|
||||
*
|
||||
* 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
|
||||
* the Free Software Foundation; either version 2.1 of the License, or
|
||||
* (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
|
||||
* GNU Lesser General Public License for more details.
|
||||
*
|
||||
* You should have received a copy of the GNU Lesser General Public License
|
||||
* along with this program; if not, see <http://www.gnu.org/licenses/>.
|
||||
*/
|
||||
/*************************************************************************/
|
||||
/* MODULE_NAME:slowpow.c */
|
||||
/* */
|
||||
/* FUNCTION:slowpow */
|
||||
/* */
|
||||
/*FILES NEEDED:mpa.h */
|
||||
/* mpa.c mpexp.c mplog.c halfulp.c */
|
||||
/* */
|
||||
/* Given two IEEE double machine numbers y,x , routine computes the */
|
||||
/* correctly rounded (to nearest) value of x^y. Result calculated by */
|
||||
/* multiplication (in halfulp.c) or if result isn't accurate enough */
|
||||
/* then routine converts x and y into multi-precision doubles and */
|
||||
/* recompute. */
|
||||
/*************************************************************************/
|
||||
|
||||
#include "mpa.h"
|
||||
#include <math_private.h>
|
||||
|
||||
void __mpexp (mp_no * x, mp_no * y, int p);
|
||||
void __mplog (mp_no * x, mp_no * y, int p);
|
||||
double ulog (double);
|
||||
double __halfulp (double x, double y);
|
||||
|
||||
double
|
||||
__slowpow (double x, double y, double z)
|
||||
{
|
||||
double res, res1;
|
||||
long double ldw, ldz, ldpp;
|
||||
static const long double ldeps = 0x4.0p-96;
|
||||
|
||||
res = __halfulp (x, y); /* halfulp() returns -10 or x^y */
|
||||
if (res >= 0)
|
||||
return res; /* if result was really computed by halfulp */
|
||||
/* else, if result was not really computed by halfulp */
|
||||
|
||||
/* Compute pow as long double, 106 bits */
|
||||
ldz = __ieee754_logl ((long double) x);
|
||||
ldw = (long double) y *ldz;
|
||||
ldpp = __ieee754_expl (ldw);
|
||||
res = (double) (ldpp + ldeps);
|
||||
res1 = (double) (ldpp - ldeps);
|
||||
|
||||
if (res != res1) /* if result still not accurate enough */
|
||||
{ /* use mpa for higher precision. */
|
||||
mp_no mpx, mpy, mpz, mpw, mpp, mpr, mpr1;
|
||||
static const mp_no eps = { -3, {1.0, 4.0} };
|
||||
int p;
|
||||
|
||||
p = 10; /* p=precision 240 bits */
|
||||
__dbl_mp (x, &mpx, p);
|
||||
__dbl_mp (y, &mpy, p);
|
||||
__dbl_mp (z, &mpz, p);
|
||||
__mplog (&mpx, &mpz, p); /* log(x) = z */
|
||||
__mul (&mpy, &mpz, &mpw, p); /* y * z =w */
|
||||
__mpexp (&mpw, &mpp, p); /* e^w =pp */
|
||||
__add (&mpp, &eps, &mpr, p); /* pp+eps =r */
|
||||
__mp_dbl (&mpr, &res, p);
|
||||
__sub (&mpp, &eps, &mpr1, p); /* pp -eps =r1 */
|
||||
__mp_dbl (&mpr1, &res1, p); /* converting into double precision */
|
||||
if (res == res1)
|
||||
return res;
|
||||
|
||||
/* if we get here result wasn't calculated exactly, continue for
|
||||
more exact calculation using 768 bits. */
|
||||
p = 32;
|
||||
__dbl_mp (x, &mpx, p);
|
||||
__dbl_mp (y, &mpy, p);
|
||||
__dbl_mp (z, &mpz, p);
|
||||
__mplog (&mpx, &mpz, p); /* log(c)=z */
|
||||
__mul (&mpy, &mpz, &mpw, p); /* y*z =w */
|
||||
__mpexp (&mpw, &mpp, p); /* e^w=pp */
|
||||
__mp_dbl (&mpp, &res, p); /* converting into double precision */
|
||||
}
|
||||
return res;
|
||||
}
|
||||
@ -2,4 +2,5 @@
|
||||
|
||||
ifeq ($(subdir),math)
|
||||
CFLAGS-mpa.c += --param max-unroll-times=4 -funroll-loops -fpeel-loops
|
||||
CPPFLAGS-slowpow.c += -DUSE_LONG_DOUBLE_FOR_MP=1
|
||||
endif
|
||||
|
||||
@ -1,93 +0,0 @@
|
||||
/*
|
||||
* IBM Accurate Mathematical Library
|
||||
* written by International Business Machines Corp.
|
||||
* Copyright (C) 2001-2013 Free Software Foundation, Inc.
|
||||
*
|
||||
* 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
|
||||
* the Free Software Foundation; either version 2.1 of the License, or
|
||||
* (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
|
||||
* GNU Lesser General Public License for more details.
|
||||
*
|
||||
* You should have received a copy of the GNU Lesser General Public License
|
||||
* along with this program; if not, see <http://www.gnu.org/licenses/>.
|
||||
*/
|
||||
/*************************************************************************/
|
||||
/* MODULE_NAME:slowpow.c */
|
||||
/* */
|
||||
/* FUNCTION:slowpow */
|
||||
/* */
|
||||
/*FILES NEEDED:mpa.h */
|
||||
/* mpa.c mpexp.c mplog.c halfulp.c */
|
||||
/* */
|
||||
/* Given two IEEE double machine numbers y,x , routine computes the */
|
||||
/* correctly rounded (to nearest) value of x^y. Result calculated by */
|
||||
/* multiplication (in halfulp.c) or if result isn't accurate enough */
|
||||
/* then routine converts x and y into multi-precision doubles and */
|
||||
/* recompute. */
|
||||
/*************************************************************************/
|
||||
|
||||
#include "mpa.h"
|
||||
#include <math_private.h>
|
||||
|
||||
void __mpexp (mp_no * x, mp_no * y, int p);
|
||||
void __mplog (mp_no * x, mp_no * y, int p);
|
||||
double ulog (double);
|
||||
double __halfulp (double x, double y);
|
||||
|
||||
double
|
||||
__slowpow (double x, double y, double z)
|
||||
{
|
||||
double res, res1;
|
||||
long double ldw, ldz, ldpp;
|
||||
static const long double ldeps = 0x4.0p-96;
|
||||
|
||||
res = __halfulp (x, y); /* halfulp() returns -10 or x^y */
|
||||
if (res >= 0)
|
||||
return res; /* if result was really computed by halfulp */
|
||||
/* else, if result was not really computed by halfulp */
|
||||
|
||||
/* Compute pow as long double, 106 bits */
|
||||
ldz = __ieee754_logl ((long double) x);
|
||||
ldw = (long double) y *ldz;
|
||||
ldpp = __ieee754_expl (ldw);
|
||||
res = (double) (ldpp + ldeps);
|
||||
res1 = (double) (ldpp - ldeps);
|
||||
|
||||
if (res != res1) /* if result still not accurate enough */
|
||||
{ /* use mpa for higher precision. */
|
||||
mp_no mpx, mpy, mpz, mpw, mpp, mpr, mpr1;
|
||||
static const mp_no eps = { -3, {1.0, 4.0} };
|
||||
int p;
|
||||
|
||||
p = 10; /* p=precision 240 bits */
|
||||
__dbl_mp (x, &mpx, p);
|
||||
__dbl_mp (y, &mpy, p);
|
||||
__dbl_mp (z, &mpz, p);
|
||||
__mplog (&mpx, &mpz, p); /* log(x) = z */
|
||||
__mul (&mpy, &mpz, &mpw, p); /* y * z =w */
|
||||
__mpexp (&mpw, &mpp, p); /* e^w =pp */
|
||||
__add (&mpp, &eps, &mpr, p); /* pp+eps =r */
|
||||
__mp_dbl (&mpr, &res, p);
|
||||
__sub (&mpp, &eps, &mpr1, p); /* pp -eps =r1 */
|
||||
__mp_dbl (&mpr1, &res1, p); /* converting into double precision */
|
||||
if (res == res1)
|
||||
return res;
|
||||
|
||||
/* if we get here result wasn't calculated exactly, continue for
|
||||
more exact calculation using 768 bits. */
|
||||
p = 32;
|
||||
__dbl_mp (x, &mpx, p);
|
||||
__dbl_mp (y, &mpy, p);
|
||||
__dbl_mp (z, &mpz, p);
|
||||
__mplog (&mpx, &mpz, p); /* log(c)=z */
|
||||
__mul (&mpy, &mpz, &mpw, p); /* y*z =w */
|
||||
__mpexp (&mpw, &mpp, p); /* e^w=pp */
|
||||
__mp_dbl (&mpp, &res, p); /* converting into double precision */
|
||||
}
|
||||
return res;
|
||||
}
|
||||
Loading…
x
Reference in New Issue
Block a user