Systolic design of frequency-domain block LMS adaptive digital filters

Abstract

Adaptive filtering in the frequency domain can be achieved by Fourier transformation of the input signal and independent weighting of the contents of each frequency bin. In certain applications, filtering in the frequency domain results in great improvements in convergence rate over the conventional time-domain adaptive filtering. In this paper, the use of word-level systolic arrays to implement frequency-domain adaptive filters based on the complex least mean square (LMS) algorithm is described. The transform employed is the discrete Fourier transform (DFT). The proposed architecture operates on a block-by-block basis and makes use of the parallelism inherent in the computational problem under consideration. The input and output data flow sequentially and continuously into and out of the systolic arrays at the system clock rate. During each clock period, processing elements of three different types operate in parallel. The most computationally demanding among them performs only three consecutive multiplications and two addition/subtractions per clock period thereby allowing a very high throughput and very fast block signal processing to be achieved at the expense of a delay of 2 L+1 samples between the input and the output, L being the block size.