Accelerating Smoothing of Velocity Perturbations in Seismic Tomography

In order to understand how a region developed geologically, we must first determine the Earth's (velocity) structure. First-arrival traveltime seismic tomography produces crustal velocity models using the results of either passive (earthquakes) or controlled-source (explosions) experiments. We use and modify a nonlinear first-arrival traveltime algorithm, which has two steps: 1) the computation of discrete first arrivals throughout a velocity model of the surveyed region, and 2) the inversion of traveltime residuals. A major limitation of the tomographic algorithm is the excessively long time required for the computation of each iterative model. The algorithm smoothes the velocity perturbations in each iteration for stabilization, improved convergence speed, and/or greater resolution of the model being computed. The smoothing algorithm, which is a mean filter with a progressively reduced window size, is the computational performance bottleneck of the tomography algorithm, as it represents about 80% of total execution time. We develop an approach to speed up the smoothing of velocity perturbations using a dual strategy based on smoothing window caching and parallel processing mapped to a graphics processing unit (GPU) architecture. With a large number of parallelizable filtering computations needed for smoothing, a GPU is the ideal target architecture as it runs multiple concurrent threads, keeps throughput high, and hides data access latency. Our GPU-based caching algorithm provides a 92x speedup over the sequential non-caching implementation.