Improving Fixed-Point Implementation of QR Decomposition by Rounding-to-Nearest

Abstract

QR decomposition is a key operation in many current communication systems. This paper shows how to reduce the area of a fixed-point QR decomposition implementation based on Givens rotations by using a new number representation system. This new representation allows performing round-tonearest at the same cost of truncation. Consequently, the rounding errors of the results are halved, which allows it to reduce the word-length by one bit. This reduction positively impacts on the area, delay and power consumption of the design.