A generalized matrix-decomposition processor for joint MIMO transceiver design
Abstract
A generalized matrix-decomposition processor is designed and implemented, which supports QR decomposition (QRD), eigenvalue decomposition (EVD), and geometric-mean decomposition (GMD), to accelerate computations in MIMO precoding/beamforming systems. The processor adopts memory-based architecture with 16 processing elements (PEs) each consisting of one CORDIC module. An improved GMD algorithm is proposed, which reduces 13.2% complexity and can be implemented by homogeneous CORDIC operations. The EVD adopts the Rayleigh quotient shift and deflation technique to accelerate convergence. The basis computations can be accomplished by mirrored operations during channel matrix decomposition. From the implementation results, the generalized processor achieves decomposition throughput of 10M, 0.99M, 2.96M matrixes per second for 4 × 4 complex QRD, EVD and GMD.
BibTeX
@inproceedings{icassp2017_ageneralizedmatr,
title = {A generalized matrix-decomposition processor for joint MIMO transceiver design},
author = {Yu-Chi Wu and Pei-Yun Tsai},
booktitle = {ICASSP 2017},
year = {2017}
}