在多圖形處理器系統上實現高度平行化之單視角立體影像轉換演算法

Abstract

立體的內容成為現代娛樂系統的主流，不僅在公開的環境，隨著3D顯示器技術的進步，也開始進入個人、行動以及家庭中。然而，現在包含立體資訊的內容的數量仍然不足以跟上3D顯示器技術的進步。單視角立體影像轉換分析現存的2D內容並加上深度的資訊使得可以撥放出立體的效果。然而，單視角立體影像轉換是需要龐大的計算量，特別是現在已有大量的2D影片仍有待我們轉換。一個快速且有效率的單視角立體轉換的實作方法是高度渴望的。本論文發表一個將高品質的單視角立體轉換演算法充分加速在多圖形處理器上的設計方法。本論文先充分分析整個多視角立體轉換演算法的演算法流程並找出含有大量平行化的部分。藉由利用多-中央處理器-多-圖形處理器系統的運算資源，使得生產量提升到95.5倍。本論文發表一個有系統化的流程將單視角立體轉換演算法平行化在多-中央處理器-多-圖形處理器系統上。發表的流程可以使得設計者充分利用不同數量的中央處理器與圖形處理器且避免資源爭奪的問題。

Stereoscopic contents are becoming the main stream for modern entertainment systems. Not only being popular in the public entertainment, the maturity of 3D display technology has also penetrated into personal, mobile, and home appliances. However, the quantity of the current stereoscopic contents does not catch up with the flourish of 3D display technology. The single view conversion analyzes the existing 2D contents and adds the depth information to enable the stereoscopic effect. However, the single-view conversion algorithm is extremely computation intensive, especially when a huge amount of existing 2D media contents still need to be converted. A fast and efficient implementation of single-view conversion is highly desired. This thesis proposes a multi-GPGPU design scheme to significantly enhance the high quality single-view conversion algorithm. This thesis first performs comprehensive analysis on the complete algorithm flow of single-view conversion and identifies a significant level of parallelism inherent in the algorithm. By leveraging the computation resources of multi-CPU-multi-GPGPU systems, this thesis has demonstrated as high as 95.5X throughput enhancement. The proposed systematic flow can guide designers to fully exploit the computing capability of different numbers of CPUs and GPGPUs and effectively avoid critical resource contention.