Applying data copy to improve memory performance of general array computations

Abstract. Data copy is an important compiler optimization which dy-namically rearranges the layout of arrays by copying their elements into local buffers. Traditionally, array copy is considered expensive and has been applied only to the working sets of fully blocked computations. This paper presents an algorithm which automatically applies data copy to optimize the performance of general computations independent of block-ing. The algorithm automatically decides where to insert copy operations and which regions of arrays to copy. In addition, when specialized, it is equivalent to a general scalar replacement algorithm on arbitrary array computations. The algorithm is fully implemented and has been applied to optimize several scientific kernels. The results show that the algorithm is highly effective and that data copy can significantly improve the per-formance of scientific computations, both when combined with blocking and when applied alone without blocking.