[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Pgcc in Deb

Josip Rodin wrote:
>> >
> > I see. I'd like to stress that my original claim about 50% is perfectly
> > valid, and I think that's sufficiently large to demand some action.
> AFAIR the speed gain in a normal program is about 5% at most.

I mean *maximum* here. About 50% was something I saw at the extreme for a
3d visualization that I was developing when I used -march=i586. I don't
think it can get any better. However, I would estimate something better
than 5% on the average. Up to 10% for the pessimistic, and up to 20% for the
optimistic estimate.

> I have seen how a program can be sped up by 45%, but it was an MP3 encoder
> in which people added code for using MMX and SSE instruction sets (is that
> the right term?), and it was compiled with -m<something>=i686... so actually
> it seems it was mostly the new code that made the program faster, not
> compiler optimizations.

Well, that sort of stuff is not very convenient at the moment. It probably
gets you back into the hardcore assembly coding, or very crufty code trying
to make use of numerical libraries. As an interesting side note, anyone
read the Altivec optimizations for PPC on /.?  I guess somebody put some
automatic vectorization for PPC C/C++ compilers but I'm not that sure.
> > It by no means implies that people with true i386's won't be able to
> > run Debian, it just gives a turbo option to people with pentiums.
> AFAIR there maybe be some slowdowns even on the processors for which the
> program was no optimized.

To the worse effect, when you compile it with -march=i686 it probably won't
run on the i386 anyway. That's why it should be considered a separate arch
if it is ever done.
> >
> > I don't think it has to turn into a flamewar every time somebody brings up
> > the topic. Take it this way, I've got a Beowulf cluster to play with, and
> > I can build it!
> Well, until you do something and display the results, nothing will happen to
> change Debian (unofficial) stance on optimizing for >386 machines.

I know that nothing changes in Debian due to the collective reluctance effect
:) When I do some *work*, I will gladly display the results. Though I have no
idea how such benchmarking should be done. Are there decent benchmarking
programs as in the Windoze world?


 ++++-+++-+++-++-++-++--+---+----+----- ---  --  -  - 
 +  Eray "eXa" Ozkural                   .      .   .  . . .
 +  CS, Bilkent University, Ankara             ^  .  o   .      .
 |  mail: erayo@cs.bilkent.edu.tr                .  ^  .   .

Reply to: