Têmpera simulada aplicada no mapeamento tecnológico de FPGAs baseadas em LUTs

Abstract

Currently there are several techniques for integrated circuit’s atribute optimization. The current focus of these techniques is to minimize the area of the given circuit. These current techniques, however, have several stages that need improvement, including the Technology Mapping stage. The technology mapping is a crucial step in the logic synthesis process, because it de?nes which set of logic elements will be used to implement the circuit in the target technology. In the literature there are several different approaches to optimize the mapping stage and currently iterative methodologies are becoming popular. This dissertation proposes a new approach to Technology Mapping of Field Programmable Gate Arrays (FPGAs), based on optimization techniques using Arti?cial Intelligence (AI), more speci?cally the Simulated Annealing technique, in order to propose an alternative solution to the problem. The utilization of an AI technique in technology mapping is promissing approach because it strongly differs from existing techniques due to the randomness factors in optimization techniques based on AI. The developed approach acts on the mapping stage called coverage, creating a solution for the circuit based on the number of k-cuts needed for a complete coverage. Each k-cut can be directly related to a FPGA’s Look-Up Table (LUT), allowing the generation of a circuit equivalent to the required logic. This approach has been implemented in the FlexMap tool, which is a framework for developing Technology Mapping methods. Tests were performed in 85 benchmarks of the ISCAS85 and MCNC91 packages, widely known in the area and commonly used for performance testing of new approaches. The tests conducted in the implemented approach had promising results, showing that the developed technique can ?nd solutions comparable in several cases to the ABC tool, which is considered state-of-the-art in the Technology Mapping process. The results obtained by the proposed approach obtained improvements in approximately 19% of the evaluated benchmarks with K=4 and 26% with K=5 over ABC’s results.