Hybrid Hexagonal/Classical Tiling for GPUs with Chapel Iterators

Abstract

Iterative stencil computations are important in scientific computing and in the mobile domain among others. Tiling is an important method to improve computational efficiency in GPUs, enhancing data reuse and cache memory usage. Diamond tiling and hexagonal tiling both enable concurrent starts. While there are multiple combinations of different tilings, and Grosser et al. have proposed a diamond and hexagonal tiling, we will attempt to construct a combine tiling that uses Chapel parallel iterators.