A CPU/GPU Heterogeneous Hybrid Parallel Algorithm of Prestack Time Migration in Local Angle-domain

Prestack time migration (PSTM) in local angle-domain (LAD) prompts more reasonable utilizations of recorded seismic data but consumes massive amounts of memory and long computing time. To fully use the computing resource and to improve the computing efficiency, a central processing unit (CPU) and graphic processing unit (GPU) heterogeneous hybrid parallel algorithm of LAD migration is developed on GPU cluster. This algorithm realizes distributed computing on different platforms successfully and contains both GPU algorithm and CPU algorithm of LAD migration. It not only uses GPUs but also utilizes CPU cores to share the computing tasks. First, a new seismic data division method is proposed to reasonably divide large-scale data among CPU cores and GPUs. Second, a new Message Passing Interface (MPI) and Compute Unified Device Architecture (CUDA) pattern is presented to call CPU cores and GPUs at the same time. Third, the CPU/GPU heterogeneous hybrid parallel algorithm is developed and is applied into the actual data. Results demonstrate that our algorithm has the same migration profiles with the CPU algorithm and high accelerating performance. The computing time of our algorithm is the shortest, which is 235 times shorter than 20 4-core CPUs and 1.7 times shorter than that of 4GPUs.