[XviD-devel] Re: Automatic benchmark / compile

[Christoph Lampert]

On Tue, 21 Jan 2003, suxen_drol wrote:

> 
> On Sun, 19 Jan 2003 11:37:50 +0100 (CET) Christoph Lampert <chl@math.uni-bonn.de> wrote:
> 
> > > > * Is there a similar method for Windows?
> > > 
> > > Who cares? ;)
> > 
> > no further comment.
> 
> windows users have access to sh&friends through cygwin or msys(mingw).
> 
> now regardint the script, how is it gonna work?
> will u have a pool of options (-march=XX, -Oxx, -fXXX) and simply try
> every combination?

Yes. I though of two groups of flags. Those that will be combined into
every combination (like -march,-O) and those that will be done "greedy"
(if it gives improvement for one setting, it should for every other -O
as well), e.g. aligment. 

E.g. I found that xvid_gmc.c was 15-20% faster on PII/450 with  

-O2 -funroll-loops -fschedule-insns 

than e.g. with -O2 alone, also faster than with -O3 and any combination of
flags and much faster than with -Os. Other "stronger" flags of these args
performed worse, like "-funroll-all-loops" and -fschedule-insns2" 

This I wouldn't have guessed, so it was good to check once. 

Also, on Mobile PIII 1GHz it was better to use -mcpu=athlon than
-mcpu=pentium3 with gcc 3.2 ... 

gruel