With the slowdown of Moore’s law and the end of Dennard scaling, the gap between compute hardware performance and the computational demands of software applications is widening. To close this gap, we need to consider innovative optimizations at all layers of the computing stack. Towards this pursuit, I work on accuracy-aware compilers with the goal to improve the performance and energy efficiency of software programs. The compilers I build significantly improve the performance and energy usage of programs by systematically trading off small amounts of computational accuracy, often without impacting end-to-end application-level quality.

In this talk, I will discuss ApproxHPVM and ApproxCaliper. ApproxHPVM is an end-to-end compiler framework for accuracy-aware optimization of deep learning workloads running on heterogeneous edge compute platforms. ApproxHPVM is end-to-end in that it only requires high-level information from developers and automatically chooses low-level approximation techniques and knobs that tune the accuracy and performance tradeoffs. The ApproxHPVM compiler supports two different automated approximation selection techniques each with its own merits: i) ApproxTuner (PPoPP’21), and ii) ApproxHPVM IR optimizer (OOPSLA’19). ApproxTuner is an approximation tuning framework with a novel three-phase strategy that enables fast ahead-of-time tuning and efficient dynamic tuning. ApproxHPVM-IR optimizer is an error budgeting framework that quantifies and represents the error tolerance of individual operations at the IR (compiler intermediate representation) level and uses it for very fast selection of approximation knobs at install-time. I will also discuss ApproxCaliper, the first application-aware neural network optimization framework that provides significantly higher speedups compared to application-agnostic tuning methods. I will also lay out a future roadmap of my research vision that focuses on enabling accuracy-aware compilers for a wide range of emerging application domains.