Double precision stencil computations on Kepler GPUs

Graphics Processing Units (GPU) have been used extensively for accelerating parallelizable applications in general, and scientific computations in particular. Stencil based algorithms are used intensively in various research areas and represent good candidates for GPU based acceleration. Since scientific computations have high accuracy requirements, herein we focus on stencil based double precision computations. For a seven-point stencil we introduce two basic implementations, which use two-dimensional and three-dimensional thread organization respectively. Different optimization techniques lead then to a total of seven different implementations, which are evaluated for two NVIDIA Kepler GPUs. The best performance is obtained for the GTX680 card, for a kernel with two-dimensional thread organization and optimized shared memory and register usage.