Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <mwv9pj2m8z.fsf@tomate.loria.fr>
Date: Fri, 10 Jan 2020 19:35:08 +0100
From: paul zimmermann <Paul.Zimmermann@...ia.fr>
To: Szabolcs Nagy <nsz@...t70.net>
CC: Szabolcs.Nagy@....com, nd@....com, jens.gustedt@...ia.fr,
	Vincent.Lefevre@...-lyon.fr, musl@...ts.openwall.com
Subject: Re: Re: musl mathematical functions

       Dear Szabolcs,

> i think libm functions are extremely rarely used with
> non-nearest rounding mode so i think
> 
>   NR accuracy >> DR accuracy >> NR symmetry >> NR speed
>   >> DR symmetry >> DR speed
> 
> where NR is nearest rounding and DR is directed rounding.

yes this makes sense.

> and by accuracy i just mean correct behavirour with respect
> to exceptions and results (i.e. small ulp errors).

note that if directed rounding is used to implement interval
arithmetic, it is very important to have the return value on
the right side with respect to the exact value (at the cost
of a few ulps of accuracy).

Paul

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.