Re: [Fwd: libfreevec benchmarks]
Hi, Konstantinos.
On Sep 09 2008, Konstantinos Margaritis wrote:
> ???????? Tuesday 09 September 2008 01:00:42 ??/?? Rogério Brito ????????????:
> > This might explain my impression in a private mail to Charles Plessy
> > that my iBook G3 was quite faster under Linux than under MacOS X.
> > Apparently, MacOS X is *made* *for* altivec machines.
>
> True, Altivec is quite extensively used in MacOS X, which is not sth I
> can say for any core component of Linux, sadly. This was exactly the
> reason I started libfreevec in the first place.
Right. What I meant with my previous e-mail was to say that if your
libfreevec provides just better scheduling or better algorithms than
what is available for G3 machines, that would be *very* welcome.
> > This would be very welcome addition to glibc. Especially for the
> > "weaker" chips that don't have any vectorial part.
>
> These won't be accelerated though, only in the case that the original
> algorithm was optimised as well,
I think that you meant "was not optimised"?
> > This part about not using altivec is interesting, since I would like to,
> > say, be able to watch DVDs with my iBook and not have them drop so many
> > frames. Decoding theora videos is also quite CPU intensive... :-(
>
> Well, I can't really say how much faster decoding will be with faster libm
> functions, but 3D definitely will be faster.
Anything that makes the current situation better is very welcome.
> > > > just by choosing a different approximation method (Taylor
> > > > approximation is pretty dumb if you ask me anyway).
> >
> > What are you using, BTW? CORDIC?
>
> No, Pade approximations. CORDIC iirc, is also not that much faster than
> Taylor.
I thought that CORDIC converged faster than Taylor series and that was
the reason why it is (was?) used on scientific calculators.
> I have some relevant paper around but I can't seem to find it right
> now, but I could look it up if you really like.
I'd like to see it, if you could find it. It is a pity that the
Wikipedia article on Padé approximations is way too short (it only gives
the definition of a Padé approximation of order (m, n)).
> Anyway, the libm rewrites will be accompanied by a math paper full with proof
> and benchmarks, so I guess you will see the method used :)
I have to say that I'm a little bit more interested on the paper than on
the implementation per-se. :-)
Thanks for your work, Rogério Brito.
--
Rogério Brito : rbrito@{mackenzie,ime.usp}.br : GPG key 1024D/7C2CAEB8
http://www.ime.usp.br/~rbrito : http://meusite.mackenzie.com.br/rbrito
Projects: algorithms.berlios.de : lame.sf.net : vrms.alioth.debian.org
Reply to: