Program + Annotations Relaxation Opportunity Sites Relaxation Configurations Candidate Binaries Autotuning Heuristic Program Relaxations Analysis Analysis Log Performance and Quality MeasurementQuality

Approximate computing trades off accuracy for better performance and energy efficiency. It offers promising optimization opportunities for a wide variety of modern applications, from mobile vision to data analytics. Recent approaches to approximate computing have relied on either manual program modification, based exclusively on programmer reasoning, or opaque automatic transformations, which sacrifice programmer control. We describe ACCEPT, a comprehensive framework for approximation that balances automation with programmer guidance. It includes C/C++ type qualifiers for constraining approximation, a compiler analysis library that identifies regions of approximable code, an autotuning system that automatically chooses the best approximation strategies, and a feedback mechanism that explains how annotations can be improved for better approximation opportunities. ACCEPT automatically applies a variety of approximation techniques, including hardware acceleration, while ensuring their safety. We apply ACCEPT to nine workloads on a standard desktop, an FPGA-augmented mobile SoC, and an energy-harvesting sensor device to evaluate the annotation process. We observe average speedups of 2.3×, 4.8×, and 1.5× on the three platforms, respectively.