Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!sdd.hp.com!caen!spool.mu.edu!munnari.oz.au!bruce!monu0.cc.monash.edu.au!monu6!minyos.xx.rmit.oz.au!stan!saturn!alan
From: alan@saturn.cs.swin.OZ.AU (Alan Christiansen)
Newsgroups: comp.arch
Subject: Re: Error magnitude in i860 fp divide.
Message-ID: <1761@stan.xx.swin.oz.au>
Date: 23 Jun 91 18:36:41 GMT
References: <1991Jun21.203429.20415@oakhill.sps.mot.com>
Sender: news@stan.xx.swin.oz.au
Lines: 24

jeff@oakhill.sps.mot.com (Jeff Enderwick) writes:


>	I'm looking at an old i860 manual, sections 9.2 and 9.3. It appears
>	that they propose fp division via reciprocal approximation and
>	multiplication. The manual states that "Worst case error is about 
>	two least-significant bits", for both SP and DP. I've tried doing
>	some error analysis, and I can't figure out how they came up with
>	a maximum error figure that small. 

>	Does anyone know how to show this result, or does anyone know the
>	number for the i860 applications group ?

I have not seen the i860 manual you are referring too. I have however seen
Motorollas dsp96002 manual and they perform division in a similar way.
Using an initial approximation of 1/x then progressive improvement
of this accuracy. eventually y/x is found by multiplying an accurate
version of 1/x times y. If this is what you are referring too I can explain
how the improved accuracy occurs.

Alan

>	Jeff Enderwick
>	jeff@otis.sps.mot.com