一個低面積及高能量效益之使用分散式算術的適應性有限脈衝響應濾波器

Abstract

本文提出了使用分散式算術的適應性有限脈衝響應濾波器架構。分散式算術是一種不使用乘法器而能夠實現有限脈衝響應濾波器的運算方法，主要是將可能的運算組合儲存在表中，透過查找表(LUT)的方式來進行運算。由於分散式算術會在固定數量的時脈週期下完成濾波器的運算，與濾波器的長度無關，因此對高階濾波器來說，分散式算術具有高吞吐量(high-throughput)的能力。過去許多研究都利用分散式算術高吞吐量的能力來實作適應性濾波器，雖然他們都克服了在分散式算術的適應性濾波器中需要更換表的問題，但若是更換後的表中有欄位(entry)不曾被使用的話會造成浪費，因此本文提出只有在濾波器運算下使用到得欄位才會即時去做更新數值的計算，這樣可以從這些未被使用的欄位減少更換表所需的計算量。所提之作法是依靠於用來做為查表位址的移位暫存器中的係數分佈情況。我們在UMC 90奈米的製成下使用所提出的分散式算術實作出LMS演算法的適應性濾波器，合成的結果顯示與現有使用分散式算術的適應性濾波器相比，在濾波器長度分別為16及32的情況下可以節省19.6%和23.4%的面積以及16.2%和22.4%的能量，但是在運算時間上會付出一些代價

This thesis presents an adaptive FIR filter architecture using Distributed Arithmetic (DA). DA realizes the Finite-impulse response (FIR) filter without multiplier by accessing the look-up table (LUT) which stores the possible combination of operations. Many researches have implemented adaptive filter with DA for its high-throughput capability that DA performs the filter operation in fixed clock cycles regardless of the filter length. Although they overcame the problem of LUT update in DA-based adaptive filter, the entry which not be used is wasted for LUT updating. This thesis proposes a new LUT updating which updates the entry in time when it is accessed in filter operations. It saves the operation from unused entry without updating calculation. The proposed approach relies on the distribution of coefficients which store in shift register as accessing address of DA. We implement the least-mean-square (LMS) adaptive filter base on the proposed on-demand DA with UMC 90-nm CMOS library. The synthesis result reveals that the proposed design saves 19.6% and 23.4% in area and 16.2% and 22.4% in energy with slight penalty in Timing for filter length 16 and 32, respectively, compare with the existing DA adaptive filter design.