Maximizing TTI RTM Throughput for CPU+GPU

There is a continuous interest for increased computational performance in seismic modelling and migration based on discretization of the 2-way wave equation. In many complex exploration settings the acoustic tilted transverse isotropic (TTI) approximation is preferred. Traditionally the CPU hardware architecture has been used, but during the recent years the GPU technology has been gaining popularity because of its significantly higher theoretical peak performance. However, not all problems are equally suited to take advantage of this potential and the low-level programming model is more demanding. In this work a comparison between a CPU only and a hybrid CPU and GPU implementation is presented. The starting point is a code well optimized for CPU systems by using long operators, coarse grids and cache optimization. The performance optimizations required to get good GPU performance are described. The success measure in the CPU+GPU port is migration throughput, not speedup of individual kernels. A comparison on a system with Intel Westmere CPU and nVidia Fermi GPU shows a performance improvement of 4.4x.