Presentation title

Addressing DAGs of Heterogeneous CPU-GPU Parallel Tasks Through High-Productivity Single-Source PHAST Library

Authors

Biagio Peccerillo and Sandro Bartolini

Institution(s)

Università degli Studi di Siena, Italy

Presentation type

Technical presentation

Abstract

Parallel architectures today are everywhere. Multi-core CPUs have now replaced single-core CPUs in every field, and also GPUs are spreading because of their performance and low-power capabilities. Embedded systems are no exception to this trend. Also, they are becoming heterogeneous, with a multi-core CPU and one or more GPUs often found in the same system. Thanks to this, embedded applications can now meet performance and power constraints that were impossible to meet otherwise. However, taking advantage of the underlying hardware is not for free, since programmers must explicitly code their application according to it. For this reason, it is crucial to invest some effort in designing high-productivity heterogeneous libraries and frameworks that can help programmers in their work. In particular, data-parallel and task-based models should be supported, in order to model the majority of embedded heterogeneous applications. Here we present an extension of PHAST Library, a high-productivity single-source data-parallel library that supports both multi-core CPUs and NVIDIA GPUs. We add the experimental support of DAGs of heterogeneous CPU-GPU tasks, giving programmers the possibility to choose the platform of execution of each task based on runtime decisions.

---

Additional material

• Presentation slides: [pdf]

For more details on this presentation please click the button below: