The de-dispersion transform corrects for dispersion which occurs when photons interact with electrons on their path through the interstellar medium. Dispersion has the effect of causing a frequency-dependent time delay in the photon's propagation. Specifically, lower frequency photons of the signal are observed later than their high-frequency counterparts. Among the many processing steps performed on the time-domain radio astronomy data, de-dispersion transform is one of the more computationally expensive steps.
AstroAccelerate's GPU implementation of de-dispersion is written in CUDA for NVIDIA GPUs and can process 4bit, 8bit, 16bit and 32bit inputs. Our de-dispersion is optimised to use the shared memory the fastest memory on a GPU. It uses a custom data format that allows us to use SIMD in the WORD technique to process more time-series elements per calculation.