From: Keith Bierman <Keith.Bierman@sun.com>
Date: November 30, 2004 3:26:55 PM MST
To: "'J3 Fortran'" <j3@j3-fortran.org>
Cc: Subject: (j3.2004) [Fwd: [Numeric-interest] correctly rounded LIBM
subset]
I know this isn't Fortran specific (nor even coded in Fortran) but I
thought it might be of interest to some of you.
-------- Original Message --------
Subject: [Numeric-interest] correctly rounded LIBM subset
Date: Tue, 30 Nov 2004 12:18:28 -0800 (PST)
From: Neil Toda <ntt@cubmeister.sfbay.sun.com>
To: numeric-interest@oakapple.net
Kwok-Choi Ng of Sun Microsystems' Floating Point Libraries group has
put together a correctly-rounded double-precision subset of libm
consisting of seven elementary transcendental functions. The rounding
direction can be set to any of the four rounding directions specified
by IEEE-754. The library is distributed as source with makefiles to
build it, and a set of tests to check the library's basic functioning.
Current development in the field of correctly-rounded libm functions
are
generally aimed at replacing vendor supplied libraries that are
advertised to be nearly correctly rounded, returning results with
maximum errors not much larger than .5 ULPs. These algorithms are of
course concerned with performance, both average, and for worse case
situations. For these attempts, a multi-stage approach seems to show
greatest promise:
A lower precision computation first, followed by a computation with
significantly more precision if the original step did not yield a
result
that could be judged correctly rounded. However, to set an upper limit
for the precision used, that precision must be proved sufficient to
allow correct rounding of the result: Still an active research topic.
The functions of LIBMCR are designed to attempt solutions at lower
precisions, but will continue, adding precision until the result can
be correctly rounded. Clearly, while these algorithms will return
correctly rounded results, there is no guarantee as to the maximum
computer-power/memory-load necessary to complete a calculation,
For this reason, LIBMCR is not considered a candidate for a production
level library, but rather a tool that can be used to further study/test
efforts in correctly rounding the elementary transcendental functions.
The library is available through Sun-Developer's "Whats New" portal:
http://developers.sun.com/prodtech/cc/
LIBMCR is supported by Sun's Floating Point Libraries group and
feedback is welcomed and can be sent to :
libmcr-comments-2004 AT sun DOT com
======================\|/==============================================
=== Neil T Toda -- o -- Sun Microsystems Inc | FP Lib Group ===
=== Neil.Toda@Sun.com /|\ Menlo Park, Ca | M/S MPK16-303 ===
=======================================================================
_______________________________________________
Numeric-interest mailing list
Numeric-interest@ucbtest.org
http://mailman.oakapple.net/mailman/listinfo/numeric-interest