f Programming Seismic Algorithms for GPUs

Graphics processing units (GPUs) have been shown to be capable of efficiently running computationally demanding seismic imaging algorithms. And the recent significant increase in expenditures by the petroleum industry for GPU clusters indicates these systems are cost effective. With this hurdle cleared, the adoption of GPUs is probably limited mainly by our ability to program seismic algorithms for GPUs. At Hess Corporation, we have moved the most computationally intensive parts of our seismic imaging<br>codes from CPUs to GPUs over the past few years. The effort involved has varied widely from code to code, from a cost of a man-month to nearly a man-year. Our one-way wave-equation migration for GPUs is a direct port of the computational algorithm used on CPUs. The Kirchhoff code required manual optimization of many of its components. An optimized reverse-time migration library was constructed by screening a set of automatically generated kernels. In this talk we will present the computational algorithms for these seismic imaging codes and discuss our software approaches and performance results.