Array padding in the functional language SAC

SAC is a functional array processing language that tries to combine generic, high-level program specifications with efficient runtime behavior. Being particularly designed for numerical applications, runtime performance critically depends on the effective utilization of the memory hierarchy. For many programs, however, it can be observed that the achieved performance significantly changes with small variations in the problem size. Array padding is a well-known optimization technique that adjusts the data layout of arrays in order to make better usage of caches. The paper presents art algorithm that derives a customized data layout from an array access pattern and a cache specification. Cache phenomena such as spatial and temporal reuse are taken into account as well as different cache architectures. The effectiveness is demonstrated by investigations on the runtime performance of the PDE1 benchmark on a shared memory multiprocessor.