ODISSEE - Accelerating facet-based gridding and degridding with GPUs

One of the most ambitious examples of scientific infrastructure across all disciplines is the SKA project, which aims to develop the world’s largest radio observatory in the coming decades. The enormous volumes of data that will be generated by this next-gen radio-interferometer must be processed within a guaranteed maximum time to prevent saturating the data pipelines and in particular, the radio imaging pipeline, while remaining as efficient as possible in terms of energy consumption (performance/watt). The gridding operation is a critical step within the imaging pipeline that maps the non-uniformly sampled visibilities onto a regular grid to allow the use of Fast Fourier Transforms (FFT). In order to fully exploit these next generation instruments’ capabilities efficiently while achieving high dynamic ranges, we have developed a GPU-accelerated implementation of a gridder, leveraging Baseline Dependent Averaging (BDA) for optimized data compression and capable of handling Direction Dependent Effects (DDE) through faceting, including the use of Jones matrices. The choice of GPU accelerators originates from the promise of higher peak performance and memory bandwidth, provided their parallelism is efficiently leveraged, and comes at the cost of memory transfer to and from the device. This gridder is implemented in a package usable both in C++ or Python and was integrated into the facet-based radio imaging package DDFacet, showing compatibility with existing workflows and allowing accuracy comparison with other implementations in a full pipeline. Leveraging state-of-the-art supercomputer capabilities, we evaluate the significant improvement in computational performance of this gridder on simulated SKA data at scale as well as the impact of data movements between host and device. SPIE link to publication: https://www.spiedigitallibrary.org/conference-proceedings-of-spie/14153/3104030/Accelerating-facet-based-gridding-and-degridding-with-GPUs/10.1117/12.3104030.full