Establish pipeline demonstrating a (possibly GPU) processing function

Outcomes from NZAPP: Four Cuda kernels have been prepared and onboarded at https://gitlab.com/ska-telescope/sdp/ska-gridder-nifty-cuda . Two kernels used for the Nifty gridder were adapted to also be used for the Nifty degridder, and a new kernel reverse_w_screen_to_stack used in the degridder to invert the third gridder kernel apply_w_screen_and_sum. HiPPO have integrated these kernels into RASCIL and NZAPP into the SEP Cuda imaging/calibration pipeline. Both integrations were demonstrated in System Demo 11.6 on 25 August 2021, along with verification of accuracy against the C++ Nifty gridder/degridder reference code. Outcomes from HIPPO: The GPU versions of both 3D nifty gridder and de-gridder supplied by the NZAPP team were provided with python wrappers (python wrapper for gridder was started in PI10 by the SIM) and integrated into the RASCIL imaging pipeline. The details about integration could be found in HIP-17 ticket for gridder and its related confluence page ( https://confluence.skatelescope.org/display/SE/HIP-17%3A+Integrate+the+GPU+nifty+gridder+into+Rascil ) and de-gridder in HIP-43 ( https://jira.skatelescope.org/browse/HIP-43 ). The RASCIL pipeline with GPU functions was compared with the RASCIL pipeline that used CPU functions only. The comparison did not reveal any major discrepancies ( https://confluence.skatelescope.org/display/SE/HIP-20%3A+Run+continuum+imaging+pipeline+with+GPU+gridder ). The performance with and without GPU function was also evaluated. The RASCIL pipeline is up to 2.5x faster for large images (8192x8192) when GPU gridder and degridder is used. However, a number of issues were discovered with the DASK scheduler, which does not fully support GPUs. These are discussed in ( https://confluence.skatelescope.org/display/SE/HIP-36%3A+Investigate+CUDA+errors+in+continuum+imaging+pipeline ) and in ( https://confluence.skatelescope.org/display/SE/HIP-35%3A+Investigate+the+maximum+number+of+frequency+channels+in+WAGG+workflows ); Repository for the gridder and de-gridder python wrappers: https://gitlab.com/ska-telescope/sdp/ska-gridder-nifty-cuda/-/tree/sim-874-python-wrapper Changes to the RASCIL pipeline: https://gitlab.com/ska-telescope/external/rascil/-/tree/hip-17_wagg_context The work was presented during 11.6 demo session: https://confluence.skatelescope.org/display/SE/2021-08-25+DP+ART+System+Demo+11.6 Future work: Investigate DASK options and possibilities of GPU scheduling.