Matrix engine for signal processing applications using the logarithmic number system

Abstract

An architecture design is presented for a device based upon the logarithmic number system (LNS) that is capable of performing general matrix and complex arithmetic, with features useful for DSP system-on-chip applications. A modified LNS addition/subtraction unit is employed in multiple execution units to achieve a maximum single-precision floating-point (FP) equivalent throughput of 3.2 Gflop/s at a clock frequency of 200 MHz. Each execution unit is capable of computing functions of the form (ab + cd)/sup e/ for e /spl isin/ {/spl plusmn/0.5, /spl plusmn/1, /spl plusmn/2} in a 5-stage arithmetic pipeline and returning a result every cycle, yielding a considerable per-cycle improvement over both floating- and fixed-point systems. Comparisons with existing devices and a single floating-point unit are given.