Dynamic Application Autotuning for Self-aware Approximate Computing

Abstract

The energy consumption limits the application performance in a wide range of scenarios, ranging from embedded to High-Performance Computing. To improve computation efficiency, this Chapter focuses on a software-level methodology to enhance a target application with an adaptive layer that provides self-optimization capabilities. We evaluated the benefits of dynamic autotuning in three case studies: a probabilistic time-dependent routing application from a navigation system, a molecular docking application to perform virtual-screening, and a stereo-matching application to compute the depth of a three-dimensional scene. Experimental results show how it is possible to improve computation efficiency by adapting reactively and proactively.