Error Analysis for Loop Programs using Anytime Instructions in Approximate Computing

Abstract:
Approximate Computing is a novel design paradigm sacrificing computational accuracy for gains in other non-functional properties. We present a tool that allows to automatically analyze approximated loop programs for the resulting error and range of output values. Our tool allows to model non-uniform distributions for the input variables. We further support the novel concept of Anytime Instructions: Anytime Instructions encode the number of most-significant mantissa bits to be computed in floating point operations. They are typically used to achieve execution time and energy reductions. First experiments using the tool show promising results.