Maximum-Throughput Software Pipelining

This paper presents UNRET (unrolling and retiming), a resource-constrained software pipelining approach aimed at finding a loop schedule with maximum throughput and minimum register requirements. UNRET works in two phases. First, a pipelined loop schedule with maximum throughput is found for a given set of resources. To do this, different unrolling degrees are explored in decreasing order of expected throughput. Farey’s series are used to perform such an exploration. For a given unrolling degree and an expected initiation interval, the software pipelining algorithm successively retimes the loop, obtaining different configurations. Scheduling is done for each configuration, thus performing a large exploration of the solution space. Second, the number of registers required by the schedule is reduced. This is done also in two steps, by reducing the iteration time and by rescheduling some operations in the schedule, attempting to reduce the maximum number of variables whose lifetime overlap at any cycle. The algorithm runs in polynomial time. The effectiveness of the proposed approach is shown by presenting results on well-known benchmarks. Results show that UNRET may obtain faster and better schedules than other approaches, also reducing the register requirements. Results also show that UNRET obtains optimal results in most cases.