標題: | 適用於多核心PlayStation 3平台之基於多層級管線模型的多媒體平行處理技術 Parallelizing Multimedia Applications Using Multistage Pipeline Model on PlayStation 3 |
作者: | 洪正堉 劉志尉 電子研究所 |
關鍵字: | 多媒體;H.264;多核心;程式設計;multimedia;H.264;multicore;programming |
公開日期: | 2008 |
摘要: | 未來的多媒體應用傾向採用極為複雜的演算法來處理大量資料,如高解析影像。純軟體解決方案可以因應多樣化的多媒體標準。以處理器為主的平台可透過軟體更新以跟上最新的多媒體標準,如此可以大大的降低研發費用並延長產品在市場的壽命。然而傳統的單一核心處理器架構並沒有辦法提供足夠的運算能力以滿足多媒體應用的即時運算需求。多核心處理器可以提供優異的運算能力以及彈性來處理未來高複雜度的多媒體應用。但是相較於單核心處理器只需處理單一指令流,多核心處理器的程式撰寫非常困難。多核心處理器程式撰寫非常花費時間而且容易出錯。多核心處理器為程式撰寫帶來許多挑戰,包含核心間的資料傳輸、同步化以及工作負載平衡。在本論文中,我們在PlayStation 3多核心平台上實現可滿足高解析即時需求的H.264解碼。我們採用了許多有效的策略來解決多核心處理器程式撰寫的問題。其中包括利用多層級管線模型來幫助簡化同步問題、MFC感知排程來減低資料傳輸問題,以及反覆程序搬移來幫助達成工作負載平衡。經過這些方法,可節省超過70%的傳輸處理問題,核心的使用率亦可超過85%。我們的H.264解碼器效能與原始版本相比提昇了28.13倍,該H.264解碼器每秒可以解碼超過24張的1080P高解析度影像。 Future multimedia applications tend to adopt extremely complex algorithms to process vast amount of data such as high-definition video. Software solutions are preferred, for they can rely on software patches to keep up with latest multimedia standards. The development cost can be reduced and time-in-market can be extended. However, conventional single core processors fail to meet real-time requirements. Multicore architectures provide sufficient computing power and great flexibility for tomorrow’s complex applications. However, multicore programming is far more difficult compared to conventional programming which consider only single instruction stream. Multicore programming is time-consuming and error-prone. It brings new challenges including inter-core communication, synchronization and load balancing. In this thesis, we fulfill real-time high-definition H.264 decoding on PlayStation 3 multicore platform. Several effective strategies are adopted to deal with multicore programming issues. Multistage pipeline model are utilized to simplify synchronization, MFC-aware scheduling help reduces communication overhead, while iterative task migration balance workload among processors. As a result, over 70% of communication overhead is hided; processor utilization is raised over 85%. Finally, 28.13 times performance gain is achieved compared to original JM source decoder, which can decode more than 24 1080p high-definition frame per second. |
URI: | http://140.113.39.130/cdrfb3/record/nctu/#GT009511631 http://hdl.handle.net/11536/38154 |
顯示於類別: | 畢業論文 |