Accelerating the 3D Elastic Wave Forward Modeling on GPU and MIC

摘要

The forward modeling of wave propagation is a widely-used computational method in oil and gas exploration. Its iterative stencil loops also have broad applications in scientific computing. However, the time-consuming iterative stencil loops greatly limit the exploration efficiency. In this paper, we accelerate the forward modeling on a number of different parallel architectures such as multi-core CPUs, NVIDIA Tesla GPUs (448-core Fermi and 2496 SP-cores & 832 DP-units Kepler) and the latest Intel MIC (61-core Knights Corner). For the GPU platform, we propose a 5-slice scheme to handle the problem of limited block registers and we design two parallel strategies to explore the applicabilities of Fermi and Kepler architectures; for CPUs and MIC, the SIMD vectorization scheme plays the most important role in acceleration. Although our complex stencil poses a great challenge for the optimization on GPU, we manage to achieve 4.31x and 6.03x speedups by Fermi and Kepler compared with a parallel CPU version that runs on 12 cores. The best speedup on MIC is 3.76x over the parallel CPU version. We also give a detailed comparison between the GPU and the MIC architectures. Our analysis on their advantages and constraints could be served as scenarios for scientists and developers to find the suitable accelerators towards target applications.