Streaming irregular arrays

Abstract

Previous work has demonstrated that it is possible to generate eicient and highly parallel code for multicore CPUs and GPUs from combinator-based array languages for a range of applications. That work, however, has been limited to operating on lat, rectangular structures without any facilities for irregularity or nesting. In this paper, we show that even a limited form of nesting provides substantial beneits both in terms of the expressiveness of the language (increasing modularity and providing support for simple irregular structures) and the portability of the code (increasing portability across resource-constrained devices, such as GPUs). Speciically, we generalise Blelloch's lattening transformation along two lines: (1) we explicitly distinguish between deinitely regular and potentially irregular computations; and (2) we handle multidimensional arrays. We demonstrate the utility of this generalisation by an extension of the embedded array language Accelerate to include irregular streams of multidimensional arrays. We discuss code generation, optimisation, and irregular stream scheduling as well as a range of benchmarks on both multicore CPUs and GPUs.