Prototype visibility streaming pipeline with distributed facets

HIPPO PI21: In PI21, the development of the GPU gridder continued. We have investigated other available gridders like MeerKAT spectral line GPU imager and DUCC (CPU gridder). We examined the polynomial method used to evaluate the convolution kernels in DUCC, but this has not yet proven to be faster on GPUs. Improvements continued on the ska-sdp-func GPU gridder as well. We have added a lookup table for convolution kernel evaluation that has improved performance by a factor of 1.47x. The investigation of the multi-GPU gridder has also progressed and an initial prototype has been added to the ska-sdp-func repository. The performance gain of multi-GPU is marginal at the moment, however, it allows us to produce larger images than a single-GPU implementation can. We have also created a version of the GPU gridder that grids visibilities into a sub-grid and does not produce an image. This is to support the PANDO team, who are planning to implement an imager based on SwiFTly. Modified version of the GPU gridder was added into the SKA SDP processing function library ( https://gitlab.com/ska-telescope/sdp/ska-sdp-func ; https://jira.skatelescope.org/browse/HIP-820 ). API is currently provisional and awaiting integration tests. API was discussed with FO (Peter Worthmann) and with the PANDO team, who are planning to use it in tests with a streaming imager.