Algorithmic advances for software radios

A software radio is a communication system in which most hardware is replaced by software, much in the same way as analog circuits were replaced by digital signal processors in the past 30 years. Software radios are fundamentally different from hardware radios, and new algorithmic ideas are needed to make them viable. Consider for example the IEEE 802.11b standard for wireless local area networks. This standard uses a modulation scheme known as Complementary Code Keying (CCK). The canonical maximum-likelihood CCK demodulator [6] uses an algorithm based on the fast WalshHadamard transform. (Indeed, the existence of such fast algorithm is a good reason for choosing this particular modulation scheme.) In order to produce 8 bits of output, the demodulator uses about 500 integer arithmetic operations, which amounts to 687.5E6 operations per second at the standard data rate of 11 Mbps without counting the overhead of loops, function call, pipeline stalls, and register spills. This computational load is onerous for a software implementation of the 802.11b standard. My best implementation of the CCK demodulator requires about 1375E6 cycles per second of data on a state-of-the-art Athlon 1700+ processor. This implementation was obtained by means of techniques similar to those used in FFTW [4] and it is probably very close to being optimal. On the other hand, an alternative demodulation algorithm that we developed runs about 4 times faster than the algorithm from [6], but while the new algorithm works fine in software, it would not map nicely into a hardware implementation. This paper reports some recent progress made by Vanu Inc. in developing algorithms for software radios. Specifically, in Section 1 we discuss the new CCK demodulator, and in Section 2 we discuss a new maximum-likelihood decoder for convolutional codes, which is meant to replace the celebrated Viterbi algorithm [9, 3] in our software radios. The two algorithms share the following feature: The running time is not constant, but it depends on the signal-tonoise ratio (SNR) of the input data. The lower the noise, the faster the algorithm. This behavior may at first appear to be