dolfin team mailing list archive

[Johan Hake <johan.hake@xxxxxxxxx>]

On Friday March 18 2011 03:01:23 Gennadiy Rishkin wrote:
> New question #149552 on DOLFIN:
> https://answers.launchpad.net/dolfin/+question/149552
> 
> Hi,
> 
> I've just installed FEniCs and it looks like an excellent library.  

Thanks!

> I have a few questions.

> CUDA :
> Is it possible to generate (GPU) CUDA code in general and for massively
> parallel assembly in particular? If it is, how is this done?

There are experimental code which do this. But none of it is merged into main. 
Take a look at:

  https://code.launchpad.net/~florian-rathgeber/dolfin/gpu-wrappers

and this email discussion:

  https://lists.launchpad.net/dolfin/msg04040.html

Eventhough there is a dolfin branch you also need patches for ffc and I guess 
also ufc to make this work.

> OpenMP/Threads:
> Is it possible to use OpenMP/Threads? Is it turned on by default, i,e.,
> does Dolfin automatically recognise a multicore machine? If it's not
> automatic, what would I have to do to use it?

In Python you can just set the number of threads to a nonzero value:

  parameters["num_threads"] = 4

But it is still under development. If I am not wrong is only cell assemble 
supported for now.

> Linear Algebra Backend:
> How to I change the backend to use Trilinos? I notice that it uses Petsc by
> default.

  parameters["linear_algebra_backend"] = "Epetra"

But you need to have compiled DOLFIN with trilinos support.

Johan

> Gennadiy