Sparse Tiling for Unstructured Mesh Applications in the OP2 Framework

Abstract—Unstructured meshes are widely-used in scientific computing for implementing numerical methods. In applications based on such abstraction parallelism is well-exposed, but poten-tial data reuse between different loops is, in general, lost. This is mainly due to the presence of indirect memory accesses (like A[B[i]]) that prevent many optimizations. Sparse tiling is a well-known code transformation that aims at obtaining data locality across different loops by scheduling sets of iterations determined at run-time that share some blocks of data. However, so far it has been ”manually ” applied to small benchmarks only. In this work we show our on-going research into integrating sparse tiling with the OP2 framework for unstructured mesh computations. Early experiments with the Airfoil benchmark show that sparse tiling improves the run-time performance of 20 % for both the sequential and OpenMP ver