Precise Management of Scratchpad Memories for Localising Array Accesses in Scientific Codes

Unlike desktop and server CPUs, special-purpose processors found in embedded systems and on graphics cards often do not have a cache memory which is managed automatically by hardware logic. Instead, they offer a so-called scratchpad memory which is fast like a cache but, unlike a cache, has to be managed explicitly, i.e., the burden of its efficient use is imposed on the software. We present a method for computing precisely which memory cells are reused due to temporal locality of a certain class of codes, namely codes which can be modelled in the well-known polyhedron model. We present some examples demonstrating the effectiveness of our method for scientific codes.