在異質雙核心處理器平台上研究實作DSP排程器及動態精細分工的H.264編碼器

Abstract

本論文的研究重點，是在異質雙核心平台上以視訊編碼為例，研究動態精細工作分配法。視訊編碼演算法所需的計算量非常大。過去業界解決異質雙核心的工作切割(partition)問題的方法，是以 profiling的方式來測量一個應用程式的每一個函式在RISC核心和DSP核心上運行的效率，然後再決定哪些函式要在RISC或DSP執行。因為DSP核心是為訊號處理運算而設計，使用DSP核心來運算會比RISC核心的效率好，因此往往系統最終的設計會將多媒體或通訊實體層的工作由DSP核心負責，而RISC核心負責流程控制和資料整合，這樣的做法，是屬於靜態工作分配法。可是未來的多媒體系統會同時執行多項工作，當DSP核心上實際運行的工作數量越多，由傳統靜態分工所設計出的應用程式只會一味地增加DSP核心負擔。所以要在異質雙核心平台上達到最佳效能，不能只考慮應用程式的計算特性的，還必須考慮到各核心實際運行的負載。因此本論文採用動態精細分工的方式來設計系統，讓兩個核心互相協調溝通，根據實際運行的負載動態地分配工作給各個核心。本論文並在OMAP 5912 平台上，基於這個架構實作出一個DSP排程器及H.264視訊編碼器的應用程式以驗證動態精細分工系統的效能。

Most modern embedded multimedia devices today are built upon heterogeneous multiprocessor (HMP) platforms. For such systems, a common practice for the industry is to perform static task partition during development time. However, due to the dynamic nature of new generations of multimedia devices, this method can not reach optimal performance when the runtime system state is different from the assumed static state at development time. The research direction of this thesis is to study a new fine-granularity dynamic partitioning/scheduling framework for HMP, where the partitioning/scheduling decision of a computationally intensive task is done at runtime. A DSP scheduler that can support this framework is developed in this thesis. In addition, an H.264 intra-frame encoder is implemented following this dynamic partitioning/scheduling paradigm. Experimental results show that, with minimal programming effort, the proposed system can outperform a pure GPP or DSP implementation when there is only one task, and even a statically-partitioned dual-core solution when there are multiple tasks.