Multithreading and GPU acceleration

[tenres]

Thanks for making this fantastic library available! We are using it to perform a multi-step deformable registration. However, the registration is quite slow currently, taking ~ 20 mins per case. Is there an introductory notebook or tutorial on how to accelerate the registration using multithreading and GPU versions of the functions? Thank you very much!

[thewtex]

Hello, thank you for the feedback -- it is wonderful to hear it is helping in your work.

Multi-threading should work out-of-the-box. We will soon have packages built against ITK 6 that may provide some performance improvements.

GPU distribution is currently limited -- a topic for future work.

[tenres]

Thanks for your reply. Regd. GPU support, is it correct that the core Elastix C++ library already has OpenCL support which we can use directly using Elastix command line tools, but ITKElastix doesn't yet? Perhaps SimpleElastix supports GPU functions?

[thewtex]

The Elastix C++ library does have OpenCL support, and I think you need to build from source with the right options to use it vs the GitHub Releases, which are CPU only (@N-Dekker can correct me if I am wrong on that). We used to distribute itk-elastix Python packages with OpenCL support, but we do not anymore. I am not aware of SimpleElastix GPU-enabled Python packages.