Viscoelastic Forward and Adjoint Modeling with OpenCL on Heterogeneous Clusters

Abstract

Efficient seismic modeling is more and more needed because of the advent of full waveform inversion (FWI). For real case FWI, an efficient usage of the available computer resources is paramount. With the diversity of processor architectures found today, this is not a trivial task. In this study, we investigate the use of OpenCL to take advantage of large heterogeneous clusters in the context of FWI. The main objective is to present a scalable, multi-device code for the resolution of the viscoelastic wave equation that can compute the gradient of the objective function by the adjoint state method. We present several algorithmic aspects of our program in details, with an emphasis on its different levels of parallelism. The performance of the program is shown with several tests performed on large clusters with nodes containing three types of processors: Intel CPUs, NVidia GPUs and Intel Xeon PHI. We obtain a speed-up of more than 80 when using GPUs compared to a single threaded implementation and a linear scaling when computations are divided on separate nodes. Our results show that OpenCL allows a better usage of the computing resources available using a single source code for a multitude of devices.