Graph based communication analysis for hardware/software codesign

Abstract

In this paper we present a coarse grain CDFG (Control/Data Flow Graph) model suitable for hardware/software partitioning of single processes and demonstrate how it is necessary to perform various transformations on the graph structure before partitioning in order to achieve a structure that allows for accurate estimation of communication overhead between nodes mapped to different processors. In particular, we demonstrate how various transformations of control structures can lead to a more accurate communication analysis and more efficient implementations. The purpose of the transformations is to obtain a CDFG structure that is sufficiently fine grained as to support a correct communication analysis but not more fine grained than necessary as this will increase partitioning and analysis time.