[Rtk-users] SART GPU usage is like 2~3%

[Cyril Mory]

Hi Albert,

In SART, many operations are performed on CPU even if the forward and 
back projections are the CUDA ones. The diagram in the 
SARTConeBeamReconstructionFilter documentation 
(http://www.openrtk.org/Doxygen/classrtk_1_1SARTConeBeamReconstructionFilter.html) 
shows all the sub-filters SART uses. Among these, some can be replaced 
by CUDA filters (all the sources, and the displaced detector), and some 
currently cannot (Extract, Subtract, Multiply, DivideOrZeroOut, Add, and 
RayBoxIntersection) because we have not implemented these operations on 
GPU yet.

If you are interested in helping do this, let me know, and I'll guide 
you through the code to get you started. Otherwise, you may want to try 
rtkconjugategradient, which is fully implemented in CUDA, instead of 
rtksart. It will require more iterations (try 20 at first) but is more 
stable, and may end up taking the same time.

Best,

Cyril


On 22/05/2018 13:36, Albert M wrote:
> Hi,
>
> I'm using GPU (NVIDIA TITAN Xp) to reconstruct tomosynthesis images by 
> SART method, i've no error and apparently results are fine.
> In addition GPU reconstruction it's 2x faster than SART CPU (normal) 
> reconstruction (200sec vs 400sec)
>
> But GPU usage is like 2~3%. Why is so low? How I can increase this?
>
> PD: I'm using cuda at SART setting SetBackProjectionFilter and 
> SetForwardProjectionFilter at 2.
>
>
> Can someone help me?
>
> Thank you!
>
>
> _______________________________________________
> Rtk-users mailing list
> 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/20180522/52863b38/attachment.html>