Resource-constrained pipelining based on loop transformations

In this paper a novel technique for resource-constrained loop pipelining is presented. RCLP is based on several dependence graph operations: loop unrolling, operation retiming, resource-constrained scheduling, and span reduction. All these operations are focused to nd a minimum length schedule able to be executed with a limited number of resources and thus maximizing resource utilization. The results obtained show that RCLP is superior to the existing software-pipelining-based approaches. This technique has also been evaluated with 300 randomly generated loop depedence graphs. In all cases a time-optimal schedule has been found.