Phase-aware optimization in approximate computing
Subrata Mitra, Manish Kumar Gupta, Saša Misailovíc, Saurabh Bagchi · Symposium on Code Generation and Optimization · 2017
This paper shows that many applications exhibit execution-phase-specific sensitivity towards approximation of the internal subcomputations. Therefore, approximation in certain phases can be more beneficial than others. Further, this paper presents Opprox, a novel system for application's execution-phase-aware approximation. For a user provided error budget and target input parameters, Opprox identifies different program phases and searches for profitable approximation settings for each phase of the application execution. Our evaluation with five benchmarks and four existing transformations show that our phase-aware optimization on average does 14% less work for a 5% error tolerance bound and 42% less work for a 20% tolerance bound.