優化於軟體定義無線電中圖形處理器加速球面解碼之效能

Abstract

在行動通訊標準日益複雜以及通用處理器日益強大下，使用通用程式語言及通用處理器實作通訊演算法變得可行，使軟體定義無線電成為具吸引力的高彈性解決方案。本文探討在軟體定義無線電下以圖形處理器實作的基頻處理演算法中，相較其他部分在執行時行為較不規律的多輸入多輸出偵測演算法之球面解碼，根據阿姆達爾定律可能成為系統的瓶頸。因此在實作時必須考慮到圖形處理器硬體架構對軟體優化，以有效利用圖形處理器提供之性能，針對球面解碼演算法實作中造成圖形處理器執行效率不佳的因素，我們透過實作上的優化以及提出的分歧減少方法，使得在中至低訊噪比環境時，我們的分歧減少方法比起未優化的實作在執行時間上平均提升1.6倍的效能，加上實作上的優化後比起未優化之實作整體增進2-3.5倍的效能。

Modern communication protocols are getting more complicated. With general purpose processor getting more powerful, software-defined radio offer an attractive alternative with its high flexibility. Graphics Processing Unit(GPU) enable software-defined radio to exploit its massive parallelism computing paradigm. While the majority of the baseband processing algorithm is highly parallel, the sphere decoder, however, due to its depth-first search nature, is not as regular as other components. According to the Amdahl's law, the efficiency of the software-defined LTE system implementation may be limited by the sphere decoding stage. In this thesis, we propose a preprocessing stage to significantly improve the extit{warp} execution efficiency of sphere decoding algorithm on GPU. The sphere decoder can run 1.6X faster in average in a middle to low SNR environment. Together with memory hierarchy related optimizations, the overall performance improvement to a sphere decoder has been improved by 2-3.5 times.