Automatic Streamization of Image Processing Applications

Pierre Guillou (1, 2) , Fabien Coelho (2, 1) , François Irigoin (2, 1)

1 Mines Paris - PSL (École nationale supérieure des mines de Paris)
2 CRI - Centre de Recherche en Informatique

New many-core architectures such as the Kalray MPPA-256 provide energy-efficiency and high performance for embedded systems. However, to take advantage of these opportunities, careful manual op-timizations are required. We investigate the automatic streamization of image processing applications, implemented in C on top of a dedicated API, onto this target accessed through the ΣC dataflow language. We discuss compiler and runtime design choices and their impact on perfor-mance. Our compilation techniques are implemented as source-to-source transformations in the PIPS open-source compilation framework. Ex-periments show lowest energy consumption on the Kalray MPPA target compared to other hardware targets for a range of 8 test applications.

Domaines: Compilation; Calcul parallèle, distribué et partagé [cs.DC]; Traitement des images [eess.IV]

Pierre Guillou, Fabien Coelho, François Irigoin. Automatic Streamization of Image Processing Applications. The 27th International Workshop on Languages and Compilers for Parallel Computing, Parasol, Sep 2014, Hillsboro, United States. pp 224-238, ⟨10.1007/978-3-319-17473-0_15⟩. ⟨hal-01096735⟩