Hardware pipelining of runtime-detected loops

Abstract

Dynamic partitioning is a promising technique where computations are transparently moved from a General Purpose Processor (GPP) to a coprocessor during application execution. To be effective, the mapping of computations to the coprocessor needs to consider aggressive optimizations. One of the mapping optimizations is loop pipelining, a technique extensively studied and known to allow substantial performance improvements. This paper describes a technique for pipelining Megablocks, a type of runtime loop developed for dynamic partitioning. The technique transforms the body of Megab-locks into an acyclic dataflow graph which can be fully pipelined and is based on the atomic execution of loop iterations. For a set of 9 benchmarks without memory operations, we generated pipelined hardware versions of the loops and estimate that the presented loop pipelining technique increases the average speedup of non-pipelined coprocessor accelerated designs from 1.6× to 2.2×. For a larger set of 61 benchmarks which include memory operations, the technique achieves a speedup increase from 2.5× to 5.6×.