A Model-based Synthesis Framework for the Execution of Dynamic Dataflow Actors

Abstract

Dataflow process networks (DPNs) offer a convenient model of computation (MoC) for modeling parallel behaviors. However, finally synthesizing them to efficient implementations on heterogeneous hardware platforms is still a difficult task. The open computing language (OpenCL) emerged as a common hardware abstraction for programming heterogeneous hardware devices supported by many hardware vendors. In this paper, we present a model-based synthesis framework which logically incorporates OpenCL as an operating system (OS) for synthesizing DPNs to software for parallel heterogeneous implementations. In general, the state of the art frameworks for DPN synthesis incorporate static analysis and scheduling, but are limited to static DPNs, and they only allow static mappings from models to platforms. In contrast, our framework employs a more generalized data dependent and actor-based dataflow model, and allows dynamic mappings from actors to platforms. We demonstrate by simple but concrete dynamic dataflow actors that the proposed framework is capable of handling efficiently dynamic token rates and dynamic data paths at runtime.