High speed GSGR matrix inversion algorithm with application to G. fast systems

Several modern communication systems, such as G.fast-based copper transmission or LTE-based wireless systems, benefit from MIMO techniques to achieve higher capacity. In most of these techniques, the inversion of channel matrices is required. Some channel matrices have special properties such as diagonal dominance or sparseness, which can be used to approximate their inverse and reduce the number of required computations. However, in some cases of practical interest, the matrices are less structured and approximations can lead to unacceptable results. This work presents a new algorithm for efficiently inverting dense matrices such as those required by G.fast vectoring engines supporting up to 64 lines. The proposed algorithm consists of a new approach for Squared Givens Rotations (SGR) and is suitable for Digital Signal Processor (DSP) embedded platforms, requiring fewer operations and without significant accuracy loss compared with standard SGR or Gauss elimination. The new algorithm was compared to others not only using double precision in Matlab but also on a DSP TMS320C6670 multicore platform from Texas Instruments. Results on inversion of random and actual G.fast channel matrices show that the proposed GSGR algorithm outperforms the baselines regarding speed without significant accuracy loss.