In this paper, a new two-step algorithm is introduced for power efficient implementation of the folding algorithm transformation of the LMS adaptive FIR filters. The first step handles the scheduling within the folding sets to reduce the switching activity using a greedy algorithm solution for the traveling sales person (TSP) NP-hard problem. The greedy algorithm gives a sub-optimal solution of the TSP problem with low implementation cost. For a typical word-length, and for large folding factors, the projected reduction in switching activity can be as large as 50%. HEAT tool was used to simulate the effect of this reduction using a typical Baugh-Wooley multiplier and the reduction in power consumption for a wireline equalization implementation was found to lie between 25% and 60%. The new algorithm is based on a simple breadth-first search approach and can be easily performed for one time only when the filter is geared to the steady-state mode. The second step involves optimal pipelining of the folded hardware elements for minimum power. The tradeoff between the number of pipelining levels and the power consumption is evaluated. To compensate for the LMS degradation due to pipelining, a 1-pole IIR compensator is introduced in the error loop for relaxed LMS. The IIR based relaxation outperforms the relaxed lookahead pipelining by 2-3 dB of output error. Another feature of the IIR relaxation is the `smoothing' nature of the filter, thus reducing the effective switching activities on the multiplier input. The combined effect of the two techniques can lead to power savings up to 65% compared to a normal folded structure. Simulations for the combined techniques are carried out for the digital subscriber loop channel and significant savings in power are demonstrated.
|Original language||English (US)|
|Number of pages||15|
|Journal||Journal of VLSI Signal Processing Systems for Signal, Image, and Video Technology|
|State||Published - Jul 2000|
Bibliographical noteFunding Information:
This research was supported by the National Science Foundation under grant number MIP-9258670.