Efficient Retiming of Unfolded Synchronous Dataflow Graphs

Abstract

Synchronous dataflow graphs (SDFGs) are widely used to model data driven programs, for which throughput is an important real-time requirement. Retiming and unfolding are important graph transformation techniques for performance optimization of SDFGs. Retiming improves the throughput of an SDFG by redistributing its initial tokens, while unfolding by scheduling several iterations of the graph. In this paper, we present an efficient and exact method to find feasible retiming and optimal retiming of an unfolded SDFG on its original SDFG, without converting it to its equivalent homogeneous SDFG (HSDFG) and therefore without further unfolding the HSDFG. The conversion procedures are usually time and space-consuming. We also extend two state-of-the-art retiming methods for SDFGs to deal with unfolding. We implement all the three methods and perform experiments on graphs with various structures and sizes to evaluate them thoroughly. The results show that the proposal method generally outperforms the extensions of existing retiming methods, especially for the large graphs and the graphs with complex structures.