high performance computing on graphics processing units: hgpu.org

Finite-difference depth migration based on one-way wave equation uses second-order, fourth-order, or other finite-order approximations for spatial derivatives. These finite-order approximations often lead to spatial dispersion errors and low accuracy. To avoid these errors, smaller mesh spacings are used, which results in huge increase in computation cost. In this paper, we develop a new spectral differentiation matrix method for approximating spatial derivatives. The approximation of spectral differentiation matrix is of infinite-order, and therefore avoids dispersion errors and increases accuracy by using a smaller number of grid points. Recently-developed GPU computational technology is well suited to the seismic migration based on frequency-domain wavefield-continuation methods because such approaches can be sufficiently decoupled. We present a frequency-domain wavefield-continuation algorithm based on spectral differentiation matrix, and realize GPU implementation of this new algorithm on Marmousi model.