Re: Compiling with and without -march=native: benchmarks

[Jaroslav Hajek]

On Mon, Sep 21, 2009 at 11:01 AM, Corrado <address@hidden> wrote:
> Dear Octave friends,
>
> here is the benchmark comparison between Octave 322 on Kubuntu 904 64 bit
> compiled from source with and without the --march=native flag set.
>
> It seem that with the new gcc the flag -march-native yields higher
> performances.
>
> What do you think?
>

The most relevant is the for loop benchmark. In the others, the bulk
of the work is done in libraries and -march=native brings you little
benefits. Did you include -march=native in FFLAGS?

> PS: I believe recompiling lapack and blas with -march=native may lead to even
> higher performances, but I have found lapack / blas recompilation rather
> troublesome.
>

It's not that hard, but regarding the BLAS, it makes little sense to
do so with the reference BLAS implementation. A tuned implementation
such as ATLAS or some of the non-free ones will bring you much better
speed.



-- 
RNDr. Jaroslav Hajek
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz