Accelerating Nested Data Parallelism: Preserving Regularity

Lars van den Haak, Trevor L. McDonnel, Gabriele K. Keller, Ivo Gabe de Wolff

    Onderzoeksoutput: Hoofdstuk in Boek/Rapport/CongresprocedureConferentiebijdrageAcademicpeer review

    23 Downloads (Pure)


    Irregular nested data-parallelism is a powerful programming model which enables the expression of a large class of parallel algorithms. However, it is notoriously difficult to compile such programs to efficient code for modern parallel architectures. Regular data-parallelism, on the other hand, is much easier to compile to efficient code, but too restricted to express some problems conveniently or in a manner to exploit the full parallelism. We extend the regular data-parallel programming model to allow for the parallel execution of array-level conditionals and iterations over irregular nested structures, and present two novel static analyses to optimise the code generated for these programs which reduces the costs of this more powerful irregular model. We present benchmarks to support our claim that these extensions are effective as well as feasible, as they enable to exploit the full parallelism of an important class of algorithms, and together with our optimisations lead to an improvement in absolute performance over an implementation limited to exploiting only regular parallelism.
    Originele taal-2Engels
    TitelEuro-Par 2020
    SubtitelParallel Processing - 26th International Conference on Parallel and Distributed Computing, Proceedings
    RedacteurenMaciej Malawski, Krzysztof Rzadca
    Plaats van productieCham
    Aantal pagina's17
    ISBN van elektronische versie978-3-030-57675-2
    ISBN van geprinte versie978-3-030-57674-5
    StatusGepubliceerd - 18 aug. 2020

    Publicatie series

    NaamLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    Volume12247 LNCS
    ISSN van geprinte versie0302-9743
    ISSN van elektronische versie1611-3349


    Duik in de onderzoeksthema's van 'Accelerating Nested Data Parallelism: Preserving Regularity'. Samen vormen ze een unieke vingerafdruk.

    Citeer dit