High-level synthesis with variable-latency components

Abstract

Listen

Translate article

This paper presents techniques to integrate the use of variable latency units in a high-level synthesis design methodology. Components used as building blocks (e.g., functional units) in conventional high-level synthesis techniques are assumed to have fixed latency values. Variable latency units exhibit the property that the number of cycles taken to compute their outputs varies depending on the input values. While variable latency units offer potential for performance improvement, we demonstrate that realization of this potential requires that high-level synthesis be adapted suitably (sub-optimal use of variable latency units can lead to performance degradation, or unnecessarily high area overheads). Our techniques to incorporate variable latency units into high-level synthesis ensure that the performance improvement is maximized, while minimizing area overheads or satisfying resource constraints. These techniques do not assume specific high-level synthesis tools/algorithms, and can be plugged in to any generic high-level synthesis system. Since area overheads may still be incurred due to the use of variable latency units, we present a novel technique, based on the concept of reduced variable latency units, to further reduce area overheads. Reduced variable latency-units only implement the low latency case behavior of complete variable latency units. The use of reduced variable latency units significantly reduces area overheads, and frequently results in RTL implementations with simultaneous area and performance improvements compared to fixed latency implementations. Experimental results show that designs optimized using the proposed techniques achieve significant performance improvements (upto 1.6 X) over designs synthesized by a state-of-the-art high level synthesis tool, frequently with simultaneous improvements in area (upto 17.9%). In addition, while we do not explicitly target power reduction, we found the variable latency optimized designs to consume 35.7% less power on the average.