Vlastnosti a implementační aspekty residuální aritmetiky pro hardwarový řešič soustav lineárních rovnic

Abstract

This dissertation thesis focuses on implementation aspect of hardware-based error-free solving of systems of linear equations. Error-free solution of linear systems is often needed in case of large, dense and ill-conditioned systems, where rounding errors can lead to long run times due to stability problems, or even hinder the solution completely. We explored the modular arithmetic approach using the Residue Number System (RNS). We have analyzed and implemented several architectures for modular multiplication and modular inverse, which are needed to implement the elimination algorithm for solving the linear systems. We have redesigned the architecture of a residual processor for solving systems of linear congruences. This is a part of a modular system for solving systems of linear equations using the residue number system. We have analyzed the implementation results in FPGA and ASIC platforms for different parameters such as word length or matrix dimension. The quality of hardware implemented algorithms was measured using the metrics of time and area and also the time-area product. Our analysis is will serve as a base for improvement of the modular system for solving systems of linear equations. The resulting system architecture permits error-free solution of dense systems of linear equations of sizes of more than 1000 equations in reasonable configuration in a few seconds using contemporary technology.This dissertation thesis focuses on implementation aspect of hardware-based error-free solving of systems of linear equations. Error-free solution of linear systems is often needed in case of large, dense and ill-conditioned systems, where rounding errors can lead to long run times due to stability problems, or even hinder the solution completely. We explored the modular arithmetic approach using the Residue Number System (RNS). We have analyzed and implemented several architectures for modular multiplication and modular inverse, which are needed to implement the elimination algorithm for solving the linear systems. We have redesigned the architecture of a residual processor for solving systems of linear congruences. This is a part of a modular system for solving systems of linear equations using the residue number system. We have analyzed the implementation results in FPGA and ASIC platforms for different parameters such as word length or matrix dimension. The quality of hardware implemented algorithms was measured using the metrics of time and area and also the time-area product. Our analysis is will serve as a base for improvement of the modular system for solving systems of linear equations. The resulting system architecture permits error-free solution of dense systems of linear equations of sizes of more than 1000 equations in reasonable configuration in a few seconds using contemporary technology.