運用平行架構及無競爭式交錯器之渦輪碼解碼器

Abstract

本論文探討使用平行架構及無競爭式交錯器之渦輪碼解碼器來達到高速的資料輸出量。我們首先分析傳統平行架構的方法，整理各個方法的優缺點，也列出影響資料輸出量的關鍵因素。文中的討論主要是針對採用多個soft-in soft-out (SISO)處理器來對單一個接收到的字碼進行解碼這種技術所衍生出的議題。除了增加SISO處理器的數目之外，我們也結合了其他平行架構的方法來大幅提昇速度。然而，該方法將會導致相當高的硬體複雜度及降低處理器的運算效能。為了解決複雜度的問題，本論文介紹了兩種多層級的網路系統來負責平行解碼器中所有SISO處理器與記憶體之間的資料傳輸。這兩個網路系統分別支援採用inter-block permutation (IBP)交錯器及quadratic permutation polynomial (QPP)交錯器之渦輪碼解碼器。此多層級網路連結系統可以有效地降低實作時平行架構電路的繞線複雜度。至於另一個難題，則須透過調整處理器的執行程序來解決。我們提出兩套可防止資料相關性造成低運算效能的策略，並制定了各自對應的處理器執行程序。在這兩種高效能的程序中，其一是針對廣泛應用進行的設計，而另一則只能在特定的條件下使用。它們縮短了解碼流程中各個功能單元的閒置時間；因此，SISO處理器的運算效能得到了改善。

基於上述之各項技術，我們實作了四個平行架構之渦輪碼解碼器。當中有兩個採用了IBP交錯器以及對應的多層級網路系統；它們皆使用了多個SISO處理器，且每個處理器在一個時脈週期中可處理複數筆資料；其中一個還利用了廣泛用途的高效能程序，讓硬體不會進入閒置的狀態。第三個解碼器則使用QPP交錯器和第二種網路連結系統；藉由這個裝置，再加上適當的控制電路，該解碼器最多可提供八個平行SISO處理器來支援在第三代合作伙伴計劃長期演進技術規格中所函括的全部區塊長度之渦輪碼解碼。最後一個解碼器亦使用QPP交錯器和多層級網路；它比其他三個解碼器有更高的平行度；另一方面，因為符合另一個高效能程序的限制條件，它也具備最高的處理器運算效能；這個解碼器的資料輸出量可達到1.4 Gb/s。實驗結果顯示我們所提的方法能夠得到預期的成效，也使得平行架構解碼器的速度有顯著的提昇。

This dissertation investigates the turbo decoders with parallel architecture and contention-free interleaver in pursuit of high throughput with reasonable cost. The benefits and disadvantages of conventional parallel schemes are examined; then the essential factors for throughput calculation are determined. Our discussions put emphasis on using multiple soft-in soft-out (SISO) decoders for single codeword. In addition to increasing the parallelism, the hybrid of parallel schemes is further applied for more speedup. However, the methodology leads to considerable complexity and inefficiency of processor. To reduce the complexity, we develop the multi-stage networks for the parallel data transmission in the turbo decoder. Two different types of apparatus are proposed for the designs using inter-block permutation (IBP) interleaver and quadratic permutation polynomial (QPP) interleaver, respectively. They can alleviate the routing congestion in the parallel design. To overcome the other difficulty, the processing schedule must be modified. We propose two different strategies to remove the data dependency and set their corresponding highefficiency schedules. One of them is aimed for general application, whereas the other is designed for specific case. The inactive periods within the decoding flow are greatly shortened in these schedules. Hence, the efficiency of the SISO decoder can increase.

Four implemented works are presented in this dissertation. The multi-stage interconnection for IBP interleavers is applied to the first two parallel turbo decoders. Both the two designs contain multiple SISO decoders, each of which can process two or more symbols per cycle. One of them operates with the general high-efficiency schedule, and its idle time is completely removed. The third design exploits another interconnection for QPP interleavers. With such apparatus and appropriate control flow, it can use at most 8 SISO decoders to decodes all codeword blocks defined in the 3rd Generation Partnership Project Long Term Evolution standard. The remaining design also adopts QPP interleavers and the multi-stage network. It has higher parallelism than the other designs; moreover, its support of specific high-efficiency schedule results in the best efficiency. This design can achieve 1.4 Gb/s while decoding size-4096 blocks for 8 iterations. The implementation results reveal that the proposed methods work successfully in the parallel architecture and raise the throughput significantly.