Reducing Pipeline Hazards via Rotation Scheduling y

The performance of the pipelining technique is highly dependent on the data dependency between instructions. In this paper, a pipeline scheduling algorithm, called hazards minimization with rotation scheduling (HAMMRS), is proposed to reduce pipeline hazards imposed by the data dependencies of the instructions in a loop. The HAMMRS algorithm is based on an architecture sensitive graph model, pipeline dataow graph. Several important scheduling properties of our model are derived. The algorithm applies a loop pipelining technique, called rotation scheduling, to a non-optimized initial schedule table and shorten the length of the schedule table by implicitly reducing data-dependent hazards. The best position of a re-scheduling node is selected by a simple formula we derived and proved. Experimental results, from several pipeline architectures, such as those having di erent pipeline types, and using or not using forwarding techniques exhibit the e ciency of the method, and show that each schedule length of benchmarks can be signi cantly reduced.