Data-Parallel Flattening by Expansion at ARRAY 2019
22 June 2019
Our paper Data-Parallel Flattening by Expansion, with Troels Henriksen and Niels G. W. Serup, was accepted for presentation at ARRAY 2019 in Phoenix on June 22.
Irregular parallel problems can be awkward to map onto GPUs, especially when different inputs produce different amounts of work. We introduce a reusable higher-order function that expands such work into flat parallel operations, hiding the details of segmented processing inside a library. Implemented in Futhark, the technique applies to a range of irregular problems and produces efficient GPU code without requiring programmers to write their own flattening machinery.