Using a Dynamic Schedule to Increase the Performance of Tiling in Stencil Computations.
Michael Freitag · 2014
A stencil computation determines the values of points in a grid of some dimensionality by repeatedly evaluating a given function of a grid point and its neighbors. The parallelization and optimization of stencil computations are subject of ongoing research. The most prevalent approach is the subdivision of the iteration domain into smaller pieces, called tiles. We give an overview of a method to increase the performance of one such tiling algorithm further by employing a dynamic schedule for tile processing, improving both load balance and cache efficiency. A set of onedimensional stencil benchmarks exhibits a performance increase of up to 20% in comparison to the Pochoir stencil compiler.