Profiling and Autotuning for Energy-Aware Approximate Programming

One promising approach to energy-efficient computation, approximate computing, trades off output precision for gains in energy efficiency. Many applications can easily tolerate small errors, especially if they are handled in a disciplined manner. However, approximation introduces an inherent tradeoff between quality of result and energy efficiency. Existing approaches lack ways to quantify and study these tradeoffs. This paper proposes tools to prototype, profile, and automatically tune the quality of programs designed to run on future approximate hardware. We describe the software layers required in such a system and discuss design considerations. We built an OCaml-based prototype of our set of tools and performed three case studies.