An Optimized GPU Kernel for Real-Time Radio Imaging

The E-Field Parallel Imaging Correlator (EPIC) is an algorithm designed to perform imaging at several frequencies using radio telescopes on much faster time scales
$O(N\log N)$
than their traditional counterparts
$O(N^{2})$
. This paper describes the new GPU kernel developed for a real-time EPIC-based imaging pipeline. Unlike the slower global memory used by the previous implementation, the kernel uses a new memory layout to store all intermediate products of the imaging sequence in on-chip memory to minimize time spent on memory transfers. Auxiliary data are stored in the shared memory for faster access. The imaging data is stored with 16-bit precision in the thread registers to reduce register pressure and simultaneously process two images. Each thread block processes one frequency, providing horizontal scalability. The optimized kernel processes 50000 all-sky images per second per frequency in real time. It is being deployed on the Long Wavelength Array in Sevilleta, New Mexico (LWA-SV).