Dependency-Aware Clustering for Variable-Grained Hardware-Software Partitioning

The increasing adoption of FPGA-based systems, calls for efficient and effective partitioning of application components between the hardware and software of the FPGA platform. In this work, we propose a technique for application-specific data dependency-aware clustering that facilitates variable-grained hardware-software partitioning. The variable granularity makes the approach suitable for both large and small applications as well as stringent resource constraints and mitigates the impact of relaxed communication models in partitioning heuristics. Validated on applications from the CHStone benchmark suite the technique achieves 15% and 7% performance improvement compared to function and basic block level approaches respectively.