CUDA
69 views
Skip to first unread message
Jens
Jul 2, 2009, 4:31:25 AM7/2/09
to FDS and Smokeview Discussions, ma...@belaschk.de
Hi,
Earlier this year, there were a discussion about implementing CUDA in
FDS. http://groups.google.com/group/fds-smv/browse_thread/thread/889bf25eecb15c0a/300b751549333742?lnk=gst&q=CUDA#300b751549333742
Is there any status for the project - and will it work with MPI?
Looking forward to hear more - think is a really great step forward,
if the project succeed.
Regards
Jens
Earlier this year, there were a discussion about implementing CUDA in
FDS. http://groups.google.com/group/fds-smv/browse_thread/thread/889bf25eecb15c0a/300b751549333742?lnk=gst&q=CUDA#300b751549333742
Is there any status for the project - and will it work with MPI?
Looking forward to hear more - think is a really great step forward,
if the project succeed.
Regards
Jens
F-Sim.de
Jul 2, 2009, 6:30:29 AM7/2/09
to FDS and Smokeview Discussions
I do not know about qFDS, but this could be interesting:
still very special and for sure not perfect, but ...
http://www.pgroup.com/resources/accel.htm
This should make it possible in the future to use GPUs more easily.
On 2 Jul., 10:31, Jens <kast...@gmail.com> wrote:
> Hi,
>
> Earlier this year, there were a discussion about implementing CUDA in
> FDS.http://groups.google.com/group/fds-smv/browse_thread/thread/889bf25ee...
still very special and for sure not perfect, but ...
http://www.pgroup.com/resources/accel.htm
This should make it possible in the future to use GPUs more easily.
On 2 Jul., 10:31, Jens <kast...@gmail.com> wrote:
> Hi,
>
> Earlier this year, there were a discussion about implementing CUDA in
Jens
Jul 2, 2009, 6:37:37 AM7/2/09
to FDS and Smokeview Discussions
Sounds really interesting.
The way I understand it, you compile the program (ie. FDS), and with
specifications to the compiler, it transfer some of the loops to the
GPU ?!?
It is way easier than the other project I linked to - but proberbly
not that efficient.
The way I understand it, you compile the program (ie. FDS), and with
specifications to the compiler, it transfer some of the loops to the
GPU ?!?
It is way easier than the other project I linked to - but proberbly
not that efficient.
Kevin
Jul 2, 2009, 8:11:16 AM7/2/09
to FDS and Smokeview Discussions
Christian Rogsch is currently implementing OpenMP in FDS. We want to
see how well this works before exploring other parallel options.
> > > Jens- Hide quoted text -
>
> - Show quoted text -
see how well this works before exploring other parallel options.
>
> - Show quoted text -
F-Sim.de
Jul 2, 2009, 2:52:14 PM7/2/09
to FDS and Smokeview Discussions
> Sounds really interesting.
It is.
> The way I understand it, you compile the program (ie. FDS), and with
> specifications to the compiler, it transfer some of the loops to the
> GPU ?!?
Yes. You need to tag the code lines you want to be done on a GPU.
Compiling the source with special compiler directives creates an
executable using GPU. I do not expect this to be the fastest possible
code and right now it is limited to Linux and NVidia GPUs but it is an
interesting peek into the future. The FDS code does not have to be
changed in detail to offer a special GPU-version but lets see.
> It is way easier than the other project I linked to - but probably
> not that efficient.
I think further improvements in code optimization and a wider spectrum
of usage will follow...
It is.
> The way I understand it, you compile the program (ie. FDS), and with
> specifications to the compiler, it transfer some of the loops to the
> GPU ?!?
Compiling the source with special compiler directives creates an
executable using GPU. I do not expect this to be the fastest possible
code and right now it is limited to Linux and NVidia GPUs but it is an
interesting peek into the future. The FDS code does not have to be
changed in detail to offer a special GPU-version but lets see.
> It is way easier than the other project I linked to - but probably
> not that efficient.
I think further improvements in code optimization and a wider spectrum
of usage will follow...
Reply all
Reply to author
Forward
0 new messages