Model-Driven Auto-Tuning of Stencil Computations on GPUs

Stencil computations are a class of algorithms which perform nearest-neighbor computation, often on a multi-dimensional grid. This type of calculation forms the basis for computer simulations across almost every field of science. The increasing computational speed of graphics processing units (GPUs) make their use for stencil computations an interesting goal. However, achieving highly efficient implementations is often nontrivial, as numerous publications attest. In this work, we propose an analytic performance model for stencil codes on GPUs, which both delivers close-to optimal performance, but at the same time does not require extensive tuning at compile or run time. We evaluate the effectiveness of our performance model using different stencil benchmarks and with various stencil radii.