DAMHSE: Programming heterogeneous MPSoCs with hardware acceleration using dataflow-based design space exploration and automated rapid prototyping

Abstract

Heterogeneous Multiprocessor Systems-on-a-Chip (MPSoCs) with programmable hardware acceleration are currently gaining market share in the embedded device domain. Largest MPSoCs combine several software processing cores with programmable logic. In these systems, reaching the optimal implementation performance is difficult because many manual and time-consuming steps are required to build, from the application specification, a prototype with measurable performances. In this paper, a method is developed that, based on state-of-the-art tools and High-Level Synthesis, deploys within less than an hour a whole hardware-software rapid prototype from a unique dataflow-based application representation: DAMHSE (DAtaflow Method for Hardware/Software Exploration). A human-driven Design Space Exploration (DSE) is conducted in order to find the most performing architectural solution, and compilable/synthesizable code is generated. The method has been tested on an image processing system with software and hardware parallelism. Results show that the obtained absolute performance (pixel/cycles) reaches state-of-the-art, and that DAMHSE leads to a heterogeneous system where performance increases significantly when the application is granted with more hardware resources. One of the greatest challenges in creating such a design automation method resides in the application behavior that may change over time and affect application concurrency and system performance. In order to overcome this problem, the design-time DAtaflow Method for Hardware/Software Exploration (DAMHSE) method is complemented with a runtime application management system that dynamically dispatches jobs (tasks) among the available processing elements (CPUs and/or hardware accelerators). Experimental results show that the performance penalty due to runtime application mapping and scheduling is limited and that the computational performance of the adaptive system remains high. Apart from the vendor-specific HLS, the tools and the frameworks used by the proposed method are open source and tutorials are available to reproduce the results.