Clustering techniques for register optimization during scheduling preprocessing
Francis Depuydt, Gert Goossens, Hillarie Man · 2002
A preprocessing technique for scheduling of large hierarchical and conditional DSP (digital signal processing) applications on a multiregister file architecture is described. Its goals are the optimization of the register cost and the reduction of the time complexity of complex techniques. These are both achieved by reducing the operation mobility through hierarchical clustering. The metric used in the clustering process is a multidimensional distance vector, based on a hierarchical and multidimensional extension of the min-flow algorithm. Network models are presented to take loop hierarchy and conditions into account. By scheduling the clusters separately, a large scheduling problem is divided into several smaller scheduling problems, which makes complex, optimal scheduling techniques feasible again for the scheduling of large graphs.>