Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!bloom-beacon!husc6!endor!siegel
From: siegel@endor.harvard.edu (Rich Siegel)
Newsgroups: comp.sys.mac.programmer
Subject: Re: questions on use of math chip by SE/30
Keywords: SE/30 68882 SANE precision
Message-ID: <1221@husc6.harvard.edu>
Date: 19 Feb 89 17:22:46 GMT
References: <3263@uhccux.uhcc.hawaii.edu> <1289@ccnysci.UUCP>
Sender: news@husc6.harvard.edu
Reply-To: siegel@endor.UUCP (Rich Siegel)
Organization: Symantec/THINK Technologies, Bedford, MA
Lines: 32

In article <1289@ccnysci.UUCP> alexis@ccnysci.UUCP (Alexis Rosen) writes:

>The '882 is identical to the '881 in terms of precision. These 'inaccuracies'
>are simply motorola's chosen precision. (I don't remember for sure, but I
>think moto used 80-bit numbers for transcendentals while Apple's SANE uses
>96 bit numbers). SANE has not changed in the /30.


	The 6888x uses a 96-bit data type to represent IEEE extended-
precision, and SANE uses an 80-bit type. However, the extra 16 bits are
set to zero, and are used, according to the 6888x manuals, "for
longword alignment and future expansion". The number of bits affecting
the precision is the same in either case: both use 80-bit calculation
precision.

	Transcendentals done in hardware are much faster, but somewhat less
accurate; a good compromise can be reached by using SANE for transcendentals
and hardware for four-banger (+ - * /) functions. If you're not paranoid
about a few digits on the end of your transcendental results, you can
use the hardware transcendentals and get about two orders of magnitude
speed increase.

		--Rich



Rich Siegel
Staff Software Developer
THINK Technologies Division, Symantec Corp.
Internet: siegel@endor.harvard.edu
UUCP: ..harvard!endor!siegel
Phone: (617) 275-4800 x305