GPU Accelerations on the 3D Elastic RTM Method

As one of the important migration methods, reverse time migration (RTM) generally accounts for a large part of the computing time. In recent decades, the desire for covering larger region and acquiring better resolution has greatly increased the algorithmic complexity of RTM. Therefore, computing platforms and optimizing methods that can better meet such challenges in seismic applications become great demands. This work focuses on accelerating the 10th-order stencil kernels from an elastic RTM algorithm by using the Nvidia GPUs. We first modify the backward process in the matrix format by adding extra layers, to generate a straightforward stencil kernel. A set of optimizing techniques including memory and computing approaches is then performed to design the RTM stencil on the K40 GPU. By further using the the streaming mechanism, we manage to obtain an communication-computation overlapping among multiple GPUs. The best performance employing four K40 GPU cards is 28 times better over an OpenMP version based on a socket with two E5-2697 CPUs.