LAUSR

Applications suited to approximation often exhibit significant value locality, both in terms of inputs as well as outcomes. In this early stage proposal - the ODIN: Outcome Driven Input Navigated approach to value locality based approximation, we hypothesize that value locality based optimizations for approximate applications should be driven by outcomes i.e., the result of the computation, but navigated with the help of inputs. An outcome-driven approach can enable computation slices, whose outcomes are deemed (approximately) redundant or derivable, to be entirely eliminated resulting in large improvements to execution efficiency. While such an approach provides large potential benefits, we address its design challenges by aiding the outcome-driven approach with input-navigation - attempting to map the value locality characteristics within inputs to that of the outcomes. To enable this, we build a novel taxonomy to categorize value locality and use it to analyze benchmarks from the PERFECT suite. We show that with oracle prediction and an ideal design, more than 80 percent of computations can be eliminated at an SNR of 17.8 or a 90 percent accuracy, thus capable of tremendous performance and energy benefits. Finally, we discuss directions towards achieving optimal benefits.