GPU-based Explicit Time Evolution Method

Jiarui Fang,Haohuan Fu,Guangwen Yang,Wei Wu,Nanxun Dai

DOI: https://doi.org/10.1190/segam2015-5856116.1

2015-01-01

Abstract:Finite-difference (FD) methods have long been among the most popular solutions for RTM and FWI. Compared with FD methods, the Explicit Time Evolution (ETE) method is able to simulate the wave propagation in acoustic media with higher spatial and temporal accuracy, at the cost of a more complicated memory access pattern. Similar to FD, ETE performs a stencil operation on every grid point, except that the coefficients of the stencil change spatially with the velocity parameter of that position. While FD methods already have highly-efficient designs on GPU platforms, in ETE, the sharp velocity discontinuities can result in un-coalesced memory access patterns. Moreover, the increased number of involved off-axis points in the stencil and the increased number of different coefficients bring more pressure for the fast buffers and memory in the GPU. To solve these issues, in this paper, we decompose the complex stencil into a number of sub-components, so as to form a better memory access pattern for coefficients and to simplify the calculation of stencil operations. Finally, we combine the decomposition scheme with 2.5D spatial and 1D temporal blocking optimizations. With one K20 GPU card, we manage to achieve 5.5x speedup compared against 12 cores Intel E5E5645 CPU.