Optimizing Stencil Computations for NVIDIA Kepler GPUs

We present a series of optimization techniques for stencil computations on NVIDIA Kepler GPUs. Stencil computations with regular grids had been ported to the older generations of NVIDIA GPUs with significant performance improvements thanks to the higher memory bandwidth than conventional CPU-only systems. However, because of the architectural changes introduced with the latest generation of the GPU architecture, Kepler, we show that existing implementation strategies used for such older GPUs are not as effective on Kepler as before. To fully exploit the potential performance of the latest generation of the GPU architecture, our implementation method uses shared memory for better data locality combined with warp specialization for higher instruction throughput. Our method achieves approximately 80% of the estimated peak performance by the roofline model, and even higher performance with temporal blocking.