high performance computing on graphics processing units: hgpu.org

Emerging architectures in embedded space are expected to make use of a diverse mix of multicores, vector-based units, GPU cores and special function accelerators. In order to facilitate mapping onto diverse architectures, different models of computation have been considered. Polyhedral Process Networks (PPNs) have been extensively used in automatic generation of task and pipeline parallel programs for embedded architectures. However, the single program multiple data (SPMD) type of data parallelism has not been addressed in the PPN model. In this paper, we propose a Data Parallel View (DPV) on PPNs which introduces abstractions necessary for capturing and exploiting data parallelism on top of the PPN model. As a proof of concept, we demonstrate how a PPN can be mapped onto a modern GPU using the DPV. By complementing the native PPN support for task and pipeline parallelism with the DPV support for data parallelism, we expect to make the best use of different types of architectural components and types of parallelism on heterogeneous architectures.