Efficient GPU-Implementation of Coherent Stacking with CUDA

Abstract

Coherent stacking is a key procedure for a class of algorithms that are used to process seismic data. The paper presents an efficient implementation of coherent stacking algorithm on CUDA-based GPUs. We discuss a complex of optimizations that allowed the implementation to reach 70% of peak hardware performance. Tests reveal linear dependency between computing time and problem size. Terabytes of seismic data cannot be placed into the memory of GPU card at once and thus the processing must be organized in portions. Optimal portion sizes where found for the following generations of Nvidia GPUs: Fermi, Kepler, Maxwell.