WSClean supports the IDG gridder, which runs on the GPU and is supposed to be a lot faster than the default gridder. We have already built the idg package on our feedstock. Unfortunately, I had to remove the idg dependency from the wsclean package because with idg, wsclean doesn't run on systems without libcuda (part of the NVIDIA graphics driver as far as I know), leading to failed tests via GitHub actions. Return code: 127. Error message: wsclean: error while loading shared libraries: libcuda.so.1: cannot open shared object file: No such file or directory
[ ] Karabo-Feedstock: add idg and libcufft dependencies to wsclean/meta.yml in the build, host, and run sections, build wsclean. libcufft might in fact be a dependency of idg, not wsclean, but I haven't tested that. Might be worth doing. I do know that wsclean doesn't work with the idg dep and without the libcufft dep.
[ ] Karabo-Pipeline: add transversal dependencies to idg and fftw3f to conda/meta.yaml and environment.yaml:
- idg =*=mpi_mpich*- fftw3f =*=mpi_mpich*
[ ] Find a way so that Karabo still runs on systems without libcuda (create wsclean-cuda and wsclean-no-cuda versions maybe?)
[ ] Test the IDG gridder on a suitable test image suite, compare generated images and runtimes
WSClean supports the IDG gridder, which runs on the GPU and is supposed to be a lot faster than the default gridder. We have already built the idg package on our feedstock. Unfortunately, I had to remove the idg dependency from the wsclean package because with idg, wsclean doesn't run on systems without libcuda (part of the NVIDIA graphics driver as far as I know), leading to failed tests via GitHub actions. Return code: 127. Error message:
wsclean: error while loading shared libraries: libcuda.so.1: cannot open shared object file: No such file or directory- idg =*=mpi_mpich*- fftw3f =*=mpi_mpich*