[RC5] Re: RC5-72 Pentium 4 optimizations

Slawek sgp at telsatgp.com.pl
Sun Sep 14 23:19:56 EDT 2003


Hi Matt,



I had took a very close look at the SSE2 and it's performance.
Not for d.net indeed. I had to optimise something for my job.


I had written a draft of how the RC5 core could use processor's resources
and I think it is in fact possible to use SSE2 in rc5 and get performance
gain
of about 10-20% (comparing to my current core).

It should also not increase heat generation (comparing to my current core).


Unfortunatelly this would mean a major rewrite of the rc5 core (completly
non-similar to organisation I've seen on other implementations).

Oh, and my current version of Pentium 4 simulator doesn't support some
things that would be required.


Currently I have not time for that.



I haven't looked at the MMX core. Maybe you can find something interesting.
It's here:
http://http.distributed.net/pub/dcti/source/archives/pub-20010416.tgz



-- 
Slawek Piotrowski

----- Original Message ----- 
From: "Matt Andrews" <andrewsm at ufl.edu>
To: <rc5 at lists.distributed.net>
Sent: Sunday, September 14, 2003 7:28 PM
Subject: [RC5] Re: RC5-72 Pentium 4 optimizations


> I was looking at the "processor comparison"
> (http://n0cgi.distributed.net/faq/cache/264.html) in the FAQ today and I
> saw that there was a Pentium MMX core.  This core apprantly performed
> better than quite a few other processors.  I was wondering exactly how
> the writer of this core got extra performance out of MMX -- unless the
> older projects were different and did not require bit rotates, it seems
> that an MMX core's need to constantly transfer back to normal registers
> for rotates would outweigh the performance gain for the parallel
> computations.  If there's still a copy of this core around, I'd be
> interested in taking a look at it, and probably Slawek would too.   If
> there was a possibility of SSE2 (which would theoretically perform twice
> as well as an MMX core) actually increasing performance on the pentium
> 4, I would join any effort to code the core.
>
> -Matt Andrews
>
>
> _______________________________________________
> rc5 mailing list
> rc5 at lists.distributed.net
> http://lists.distributed.net/mailman/listinfo/rc5
>




More information about the rc5 mailing list