Synthesis of application specific multiprocessor architectures for process networks

Abstract

In this paper, we address the problem of synthesis of application specific multiprocessor SoC architectures for process networks of streaming applications. An application is modeled as Kahn Process Network (KPN) which makes the parallelism present in the application explicit. The synthesis process involves selection of computation modules, memory modules, communication architecture and mapping of processes of KPN on compute units and channels on memory modules. Our solution minimizes hardware cost while taking into account the performance constraints. One of the salient features of our work is that it takes into account the additional overheads because of data communication conflicts. Our method uses average processing requirements of KPN to handle data dependent behavior of processes and cycles within the KPN. In contrast to others, we do not perform static scheduling, only mapping and synthesis is done.