[Trilinos-Users] trilinos and CUDA

[Chris Baker]

I agree with Matt. It should be noted that classes in the Tpetra library are
templated on scalar and ordinal types. One purpose was to allow improved
performance on platforms where smaller data types would allow it, such as
GPUs.

Chris


On Tue, Jul 28, 2009 at 12:46, Matt G <mgoodman at email.arizona.edu> wrote:

> Also notably most of the acceleration of CUDA comes from its capacity for
> fast 32 floating point operations.  If you have to do anything with
> precision higher than that, it isn't a great solution.  I would be happy to
> share the SOR Poisson solver I wrote with anyone though.  Still in
> pre-alpha, bit if it gave anyone a jump, I would be happy to give it away.
>
> -Matt
>
>
> On Tue, Jul 28, 2009 at 6:59 AM, Chris Baker <cgbaker at software.sandia.gov>wrote:
>
>> Gianluca,
>>
>> The Tpetra library of distributed linear algebra objects is currently
>> being refactored to supported multi-core/many-core shared memory nodes. The
>> plan is to support CUDA for Tpetra objects as well (rudimentary support is
>> already in place). By combining Tpetra with Trilinos's Belos linear solver
>> or Anasazi eigenvalue solver, this would provide the ability to run these
>> solvers mostly on the GPU (where "mostly" refers to the large
>> vector/multivector and matrix operations).
>>
>> This capability may release with the next version of Trilinos in
>> September. I say may because our priority is multi-core CPUs before GPGPU.
>> However, most of our GPU implementations have been tested outside of
>> Trilinos and are waiting to be moved in.
>>
>> Chris
>>
>>
>> On Tue, Jul 28, 2009 at 06:37, Gianluca Moro <giangiammy at gmail.com>wrote:
>>
>>> Hi all,
>>>
>>> I'm working on GPGPU, in particular using CUDA, in
>>> a numerical linear algebra application.
>>>
>>> As part of the work there will be the necessity of a
>>> Linear Solver Package, something like the AztecOO
>>> Trilinos package.
>>>
>>> Do you know of a similar package?
>>> If not, does anyone tried to implement it?
>>> (with or without success?)
>>>
>>> If I do not find anything, and I implement it by myself,
>>> is there interest on such a package?
>>>
>>> Thanks
>>> Gianluca Moro
>>>
>>> --
>>> Gianluca Moro                  http://www.giammy.com
>>> Ph.D. Student                   http://www.math.unipd.it/~giammy/<http://www.math.unipd.it/%7Egiammy/>
>>> Department of Pure and Applied Mathematics - University of Padova
>>>
>>> _______________________________________________
>>> Trilinos-Users mailing list
>>> Trilinos-Users at software.sandia.gov
>>> http://software.sandia.gov/mailman/listinfo/trilinos-users
>>>
>>>
>>
>> _______________________________________________
>> Trilinos-Users mailing list
>> Trilinos-Users at software.sandia.gov
>> http://software.sandia.gov/mailman/listinfo/trilinos-users
>>
>>
>
>
> --
> Matthew Goodman
>
> =====================
> Find me on LinkedIn: http://tinyurl.com/d6wlch
> Follow me on twitter: http://twitter.com/meawoppl
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20090728/ee58d990/attachment.html