Dynamically adaptive and reliable approximate computing using light-weight error analysis

Prior art in approximate computing has extensively studied computational resilience to imprecision. However, existing approaches often rely on static techniques, which potentially compromise coverage and reliability. Our approach, on the other hand, decouples error analysis of the approximate accelerator from quality analysis of the overall application. We use high-level, application-specific metrics, or Light-Weight Checks (LWCs), to gain coverage by exploiting imprecision tolerance at the application level. Unlike metrics that compare approximate solutions to exact ones, LWCs can be leveraged dynamically for error analysis and recovery. The resulting methodology adapts to output quality at runtime, providing guarantees on worst-case application-level error. To ensure platform agnosticism, these light-weight metrics are integrated directly into the application, enabling compatibility with any approximate acceleration technique. Our results present a case study of dynamic error control for inverse kinematics. Using software-based neural acceleration with LWC support, we demonstrate improvements in coverage, reliability, and overall performance.