Polyhedral Dataflow Programming: a Case Study

Abstract : Dataflow languages expose the application's potential parallelism naturally and have thus been studied and developed for the past thirty years as a solution for harnessing the increasing hardware parallelism. However, when generating code for parallel processors, current dataflow compilers only take into consideration the overall dataflow network of the application. This leaves out the potential parallelism that could be extracted from the internals of agents, typically when those include loop nests, for instance, but also potential application of intra-agent pipelining, or task splitting and rescheduling. In this work, we study the benefits of jointly using polyhedral compilation with dataflow languages. More precisely, we propose to expend the parallelization of dataflow programs by taking into account the parallelism exposed by loop nests describing the internal behavior of the program's agents. This approach is validated through the development of a prototype toolchain based on an extended version of the ΣC language. We demonstrate the benefit of this approach and the potentiality of further improvements on relevant case studies.
Type de document :
Communication dans un congrès
SBAC-PAD 2018 - 30th International Symposium on Computer Architecture and High-Performance Computing, Sep 2018, Lyon, France. IEEE, pp.1-9, 2018
Liste complète des métadonnées

Littérature citée [10 références]  Voir  Masquer  Télécharger

https://hal-cea.archives-ouvertes.fr/cea-01855997
Contributeur : Laure Gonnord <>
Soumis le : mardi 14 août 2018 - 10:09:14
Dernière modification le : mardi 13 novembre 2018 - 17:42:25
Document(s) archivé(s) le : jeudi 15 novembre 2018 - 12:45:02

Fichier

finalversion_hal_sbacpad18_FGM...
Fichiers produits par l'(les) auteur(s)

Identifiants

  • HAL Id : cea-01855997, version 2

Collections

Citation

Romain Fontaine, Laure Gonnord, Lionel Morel. Polyhedral Dataflow Programming: a Case Study. SBAC-PAD 2018 - 30th International Symposium on Computer Architecture and High-Performance Computing, Sep 2018, Lyon, France. IEEE, pp.1-9, 2018. 〈cea-01855997v2〉

Partager

Métriques

Consultations de la notice

146

Téléchargements de fichiers

68