GPU-based Acceleration of Radio Interferometry Point Source Visibility Calculations in the MEQtrees Framework

摘要

Modern radio interferometer arrays are powerful tools for obtaining high resolution, low frequency images of objects in deep space. An inverse Fourier transform of a model sky-intensity map will produce the corresponding Point Source Visibilities: the raw output of an interferometer. Simulated visibilities can be used to test models of factors affecting the accuracy of observed data, such as radio frequency interference. We describe a GPU/CUDA implementation of the Point Source Visibility calculation module within theMeqTrees software suite. For a large numbers of sources, we achieve an 18ud7 speed-up over the existing CPU module, with the parallel component running up to 120ud7 faster. With modifications to the MeqTrees memory management system to incorporate GPU memory operations, a speed-up of 24ud7 is achievable.