Deploying OpenMP on an embedded multicore accelerator

Multiprocessor systems-on-chip (MPSoC) are now considered first-class citizens both in the embedded systems and in the high-performance computing arenas, in the form of specialized or general-purpose accelerators. Programming models for such systems is currently a hot research topic, and as a general rule require deep programmer knowledge of the underlying hardware architecture. In this paper we present the implementation of OpenMP, one of the most intuitive and productive programming models, on the STHORM accelerator. This particular platform provides a shared-memory substrate which OpenMP requires. An innovative feature of our design is the deployment of the OpenMP model both at the host and the fabric sides, in a seamless way, which provides the programmer with a simple but effective interface for offloading and executing OpenMP kernels on the MPSoC. The optimized runtime environment provides full OpenMP support despite its small footprint (less than 10KB for a 16-core cluster) and can sustain close-to-ideal speedups in computationally intensive applications. We detail on design issues we faced along with their solutions, given the limited available resources.