[MEncoder-users] Fwd: Questions on Mencoder performace

[Ivo]

On Tuesday 06 June 2006 19:45, Raphael wrote:
> larrystotler at netscape.net wrote:
> > -----Original Message-----
> > From: Corey Hickey <bugfood-ml at fatooh.org>
> >
> >> I don't know specifically about your systems, but I wouldn't expect
> >
> > much performance out of a P4. I did some rough tests with lavc and
> > x264 a while back and found >that a 1.2 GHz Athlon Thunderbird was
> > outperforming a 1.7 GHz P4 by 10-20%. As far as I know early Athlon
> > processors were roughly equivalent to late P3s, >clock-for-clock, so
> > the numbers you posted don't seem out of line.
> >
> > Hmmm, I was under the impression that the P4 was actually optimized
> > for stuff like this and that it should outperform the Athlon in this
> > regard.  I've never been a fan of the P4 series, but I ended up with
> > with some chips and was hoping that they would be well suited for this
> > type of stuff.  Oh well.  Thanx for the insight.
> >
> > I appreciate all the advice that has been given to me by the list
> > members.  Thanx.
> > ___________________________________________________
>
> Whenever anybody starts being rude on this list it's because they don't
> know the answer. So don't worry about that :)
>
> Your best bet is to experiment.  There are too many unknowns and in
> particular of course the behaviour of the compiler will make a
> considerable difference.  One version of gcc with one set of
> optimisation flags may accidently squeeze some crucial piece of data
> into the L2 cache on a particular processor etc.   gcc has never given
> any guarantees and I remember compiling code with higher optimisation
> levels that ran more slowly and finding that after using -Os it didn't
> fit in cache etc.
>
> You may be able to find tools that give you useful stats (e.g.
> cachegrind http://valgrind.org/info/tools.html). If you can run
> systematic tests using such tools with as many different gccs and
> options and cpus as possible then I am sure we would all be very
> interested to see the results.

You might also try compiling with icc, the Intel C Compiler. libmpeg2 used 
to be broken when compiled with icc (version 8 I believe) but that might 
have been fixed. But libavcodec ran fine when I last tried. Intel claims 
its compiler generates better code specifically for Intel CPU's. By 
default, icc emulates gcc as far as inline assembly is concerned, so 
MMX/SSE(2) specific code should get compiled in as usual.

--Ivo