Enabling adaptive loop pipelining in high-level synthesis
Steve X. Dai, Gai Liu, Ritchie Zhao, Zhiru Zhang · 2017
Loop pipelining is an important optimization in high-level synthesis (HLS) because it allows successive loop iterations to be overlapped during execution. While current HLS pipelining approach achieves high performance for loops with regular and statically analyzable program patterns, it remains challenging to pipeline loops with irregular memory accesses, irregular dependence patterns, and unbalanced workload. The lack of support for dynamic program behaviors results in conservatively synthesized pipelines that sacrifice performance for maintaining presumed regularity. In this paper, we survey some of our recent work that addresses these challenges using a coordinated dynamic-static approach for enabling high-throughput pipelining of irregular loops. We propose to augment the HLS pipeline with dynamic scheduling to adapt to data-dependent behaviors, while employing static compile-time optimizations to minimize the hardware overhead associated with runtime optimization. Experimental results demonstrate that our proposed techniques can significantly improve effective pipeline throughput while conserving hardware resources.