[Rtk-users] Multi GPU capabilities

[vincent]

Hi Simon,

thank you for the quick reply.  I'll try the splitting strategy.

Best regards,

Vincent

On 15.10.19 17:53, Simon Rit wrote:
> Hi,
> No. This is quite a challenge to implement this and we have no 
> resources on this topic. My first attempt to do this would be to use 
> ASTRA <http://www.astra-toolbox.com/> from RTK. RTK only automagically 
> select the "best" GPU, see here 
> <https://github.com/SimonRit/RTK/blob/master/utilities/ITKCudaCommon/src/itkCudaContextManager.cxx#L67>.
> For FDK, I think it would be easy to split the volume and ask each GPU 
> to reconstruct a specific part of the volume (but I never did it and 
> RTK would need to allow parameterization of the device which it 
> currently doesn't).
> Note that we don't use the unified memory framework.
> Simon
>
> On Tue, Oct 15, 2019 at 5:43 PM vincent <vl at xris.eu 
> <mailto:vl at xris.eu>> wrote:
>
>     Hello everyone,
>
>     I was wondering if RTK automagically spread the workload over several
>     GPU's when available on a machine ?  I tried to find the answer by
>     myself, but up to now, the only information I could get were that:
>
>     - cuda provides with a unified memory framework supposed to simplify
>     memory management,
>
>     - class itkCudaUtil has members that identify all the GPU's
>     present on
>     the computer.
>
>     I had a look on the other itkCuda*** classes but found nothing that
>     could help me understand if multiple GPU's are managed by RTK.
>
>     Would someone would be so kind as to help me find an answer ?
>
>     I thank you very much in advance,
>
>     best regards,
>
>     Vincent
>
>     _______________________________________________
>     Rtk-users mailing list
>     Rtk-users at public.kitware.com <mailto:Rtk-users at public.kitware.com>
>     https://public.kitware.com/mailman/listinfo/rtk-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://public.kitware.com/pipermail/rtk-users/attachments/20191015/4ba1d031/attachment.html>