Přibližná ekvivalence pro aproximativní počítání
Project Period: 1. 1. 2016 - 31. 12. 2018
Project Type: grant
Agency: Czech Science Foundation
Program: Standardní projekty
approximate computing; genetic programming; evolvable hardware; relaxed equivalence checking; automata; logic
Approximate computing is a promising approach to obtain energy-efficient computer systems. It exploits the fact that many applications are error resilient, i.e., do not require a perfect output to be produced. An open problem is how to effectively obtain approximations that are good compromises between the error ratio, power consumption, and performance. Using evolutionary algorithms for the approximation has led to promising results, but it suffers from scalability problems in evaluating candidate solutions. For that, we propose a novel way: using advanced methods of formal verification redesigned to quickly calculate distances between candidate approximations and the reference implementation, which we call relaxed equivalence checking. The project seeks the following original contributions: (1) efficient algorithms for relaxed equivalence checking of combinational (stateless) and sequential (stateful) systems, (2) approximation algorithms based on genetic programming using the proposed relaxed equivalence checking, (3) experimental evaluation of the proposed approximation methods.