Discussion related to cado-nfs

[Paul Zimmermann <Paul.Zimmermann@inria.fr>]

       Hi,

I'm redirecting to the new list.

Do you have a concrete example in cado-nfs where "#pragma omp for simd"
is better than "#pragma omp for" ?

Paul Zimmermann

> Date: Sat, 9 Mar 2024 16:06:29 +0100
> From: Laël Cellier <lael.cellier@grenoble-inp.org>
> 
> Hello,
> 
> in a similar request, why not turn the #pragma omp for
> 
> into
> 
> #prama omp for simd
> 
> where it’s possible ? I’m noticing gcc tends to vectorize nothing…
> 
> Le 20/08/2012 à 9:29 AM, Pierrick Gaudry a écrit :
> > On Tue, Aug 07, 2012 at 11:36:31AM +0430, hamid reza arkian wrote:
> >> Hi,
> >>
> >> I tried GMP-ECM project (the branch that is currently under active
> >> development for GPU devices) on a GPU cluster and got good results.
> >> I want to replace ecm part of sieving step in CADO-NFS with GMP-ECM(that
> >> support GPU), but I have the following questions:
> >>
> >> - Basically, it's a good idea or not?
> >> - Is the GMP-ECM faster than ecm part of CADO-NFS? and if so, why the
> >> developers don't do this replacement,yet?
> >> In addition, if all things is OK and it's a good idea, it would be 
> >> pleasure
> >> if you give me some tips and points before doing that.
> >>
> >> Thanks in advance,
> >> Hamid
> > Hi,
> >
> > Using GPU for the ECM step in the so-called cofactorization step of the
> > sieving phase is something to consider. However, this is not as simple as
> > pluging gmp-ecm within Cado-nfs.
> >
> > Here are a few thoughts on the topic of ECM for NFS:
> > - the size of the numbers to be tested within NFS is small, and few
> >    curves are tested. This is not the traditionnal target of GMP-ECM, and
> >    therefore another version of ECM was written in CADO-NFS (mostly by
> >    Alex Kruppa, who is also a developper of GMP-ECM).
> > - Kruppa's thesis contains many details on this topic.
> >    http://tel.archives-ouvertes.fr/tel-00477005/en/
> > - the main question for using GPU is how to take advantage of the
> >    parallelism. I think that GPU implem of GMP-ECM currently runs many
> >    curves in parallel to make an optimal use of the card. But for the
> >    application to NFS, only few cruves are needed, and we want a
> >    parallelization at the input number level.
> > - the question of the transfers between the card and the cpu must be
> >    addressed. I tend to think that it would be interesting to accumulate a
> >    list of numbers to be tested and then send all of them at once, and
> >    keep the CPU busy with sieving while the GPU tries to finish the
> >    cofactorization. This implies some reorganization of the sieving code
> >    that is not so simple.
> >
> > Regards,
> > Pierrick
> >
> >