Acceleration of the Anisotropic PSPI Imaging Algorithm with Dataflow Engines

SUMMARY Although time-domain depth migration techniques have been successfully ported to run on modern hardware accelerators, their ultimate obstacle is the I/O overhead present during the imaging step. Frequency-domain depth migration algorithms overcome this limitation and can exploit the full potential of new computing technologies. In particular, our implementation of the Phase Shift Plus Interpolation (PSPI) method is characterized by fast running time, good quality results under low signal-to-noise ratio conditions and excellent results for steep dips. We provide a novel computational dataflow scheme to perform acceleration of PSPI on a generic dataflow engine. We present speedup results obtained on the state-of-theart dataflow technology for synthetic VTI datasets. Our measurements indicate that a dataflow approach can achieve high speedups despite larger and larger computational domains, the increased complexity of the anisotropic approach and the I/O overhead during angle-gathers calculation.