Approximating Q Propagations for Elastic Modeling on GPUs

Propagating wavefields using the explicit finite difference method is the kernel of reverse time migration (RTM) and high-end velocity algorithms in seismic applications. In recent decades there has been a significant increase of interest in the seismic exploration community to invert the image of the subsurface in larger regions and higher resolutions in the elastic media, which brings tremendous computing challenges. As a result, the optimizing methods for improving the performance of the wavefield propagation are in great demands. This work manages to boost the performance of the wavefield propagation in 3D elastic scenarios by approximating the Q propagation and using the multi-GPU platform. We first extend the constant-Q formulation from the 2D viscoelastic case to the 3D viscoelastic case. Different optimization techniques on GPUs are then described for an efficient modeling kernel. We also propose a set of schemes to reduce the computation to further improve the performance. The experimental results show that we can achieve significant performance speedups of 60 to 200 times with 4 GPUs over the CPU-based solution as a function of Q.