Custom Computing Design and Implementation for Multiple Dedispersion with GPU

摘要

Pulsar searching requires a real-time coherent de-dispersion process on an enormous stream of complex voltage data. We present a many-core accelerated de-dispersion pipeline, ACDT, which exploits the custom computing design for multiple de-dispersion on GPUs. The ACDT implementation optimizes the de-dispersion by switching to on-chip shared memory, adopting customized FFT with the overlap-save method, and overlapping the computation with transfer by a two-stage pipeline. The overall performance of ACDT is improved by 2 to 4 times when multiple DMs are processed in sequential compared to the state-of-the-art CDMT package.