Representing mapping and scheduling decisions within dataflow graphs
Christian Zebelein, Christian D. Haubelt, Joachim Falk, Tobias Schwarzer, Jürgen Teich · Forum on specification and Design Languages · 2013
Choosing the right programming model for multiprocessor System-on-Chip (MPSoC) platforms is a challenging task: in order to provide for automatic hardware/software synthesis in a model-based design flow, the system model should be architecture-independent on the one hand, but should also allow the back-annotation of architecture-dependent mapping and scheduling decisions on the other hand. Here, architecture-independent dataflow models are particularly suitable. In a dataflow model, concurrent processes (actors) communicate via packets transmitted over channels. However, the back-annotation process is difficult, and previous approaches are either restricted to static dataflow actors or static schedules. Also, the semantics of the underlying dataflow actors are often different from the semantics of the scheduling mechanism, which limits the compositionality of the dataflow model. In this paper, we propose a modeling approach which unifies the representation of dataflow actors and scheduling mechanisms, while also providing for both, dynamic dataflow actors and dynamic scheduling decisions. We describe how various scheduling schemes can be represented by our approach, and show the applicability of the proposed approach by means of synthetic dataflow graphs.