Data-Dependent Loop Approximations for Performance-Quality Driven High-Level Synthesis

Abstract

Approximate computing (AC) obtains performance or energy gains by trading off computational accuracy in error-tolerant applications. At the hardware level, AC-aware high-level synthesis tools apply operation-level approximations to generate a quality-reduced register-transfer level design from an accurate high-level description. While existing tools primarily target energy savings, we focus on performance optimizations. Loops are often the most performance-critical application code structures. In a loop, iterations can have different impacts on output quality due to an inherent data-dependency of approximations. Exploiting iteration-wise data variations, we present a novel technique that clusters iterations according to data statistics and applies different approximations in each cluster to maximize performance gains. Performance can be improved by up to 76%, with up to 21.7% stemming from clustering.