系統資料頻寬之研究

Abstract

資料頻寬問題長久以來就是嵌入式系統效能的瓶頸。若一個系統無法提供足夠的資料頻寬，或運算核心的頻寬需求過高，會造成運算核心缺乏足夠運算資料，因而使運算核心無法發揮全部的運算能力，進而影響到系統的效能。為了發揮系統的效能，本論文探討如何提高系統中可用頻寬使用率與降低頻寬需求。本論文以系統中資料傳輸的過程為主軸，針對傳輸源頭到終端所會遭遇的頻寬問題進行改善的研究。改善方法的核心概念是利用傳輸或存取資料之間的相關性，搭配系統中硬體裝置的工作特性提高頻寬使用率，或是重複利用已有資料減少頻寬需求。在資料傳輸的源頭，也就是記憶體控制器的部份，本論文探討如何透過存取排程，改善資料間的相關性，利用記憶體本身存取特性與系統匯流排協定的工作特性來提高的資料頻寬使用率。在改善資料傳輸源頭的頻寬後，瓶頸便落在系統匯流排上，因此本論文接著探討改善採用先進封包式協定匯流排頻寬使用率的方法。在資料傳輸末端的運算核心部份，本論文則是根據資料間存取的空間與時間的相關性，探討如何透過有效地資料重複使用，降低運算核心的頻寬需求。本論文研究的運算核心功能包含了視訊編解碼與早期視覺處理，這兩類運算的頻寬需求皆十分可觀。在視訊編解碼部份，本論文針對移動補償提出了一利用輸入移動向量與巨方塊種類之內容特性的運算核心硬體架構，可降低頻寬需求達72%。而在早期視覺處理部份，本論文針對平均飄移濾波演算法與立體視覺的立體匹配演算法，提出可減低頻寬需求的運算核心硬體架構。在平均飄移濾波架構中，本論文根據平均飄移向量大小的特性，提出部分更新乒乓暫存記憶體架構，可減少八成畫面記憶體頻寬需求。另一方面，在立體匹配演算法中，本論文根據演算法本身資料存取在空間與時間上的侷限性與相關性，提出部份列資料重複使用與擴張視窗減少存取兩方法，來達到大幅減少立體匹配資料頻寬需求。本論文基於資料相關性的頻寬改善方法，可提高可用頻寬與降低頻寬需求，有效改善系統中資料頻寬的問題，進而幫助提高系統整體效能。

Data bandwidth issue has long been the performance bottleneck of an embedded system. The computation cores in a system cannot maximize their utilization without enough data. This is usually a result of insufficient available data bandwidth or excessive data bandwidth requirement. Being aware of the importance of the data bandwidth issue, this dissertation addressed the bandwidth issue from the source to the destination of data transfers. The core concept was to facilitate the address and data correlation among accesses to solve the data bandwidth issue. Exploiting these correlations can increase bandwidth utilization given a device’s access characteristics and can also reduce bandwidth requirement through data reuse. In particular, this dissertation explored methods to increase the bandwidth utilization of memory controllers by taking the advantage of the characteristics of external memories and new advanced data transfer protocol. After improving the bandwidth utilization at the source of data transfers, this dissertation focused on improving the bandwidth utilization of a bus interconnect adopting advanced protocol under the traditional share-link topology. Finally, bandwidth requirement reduction techniques have been studied at the destination of data transfers. For video coding and early vision tasks, these techniques performed data reuse based on algorithm’s data access characteristics, such as the spatial ad temporal locality among data accesses. In video coding, this dissertation proposed a combined frame memory motion compensation (CFMMC) architecture that was capable of reducing the bandwidth requirement by up to 72% based on the characteristics of input motion vector and macroblock type data. In early vision tasks, this dissertation proposed a meanshift filtering architecture that used the proposed partial-update ping-pong buffer (PUPPB), which was based on the access locality due to intermediate meanshift vector characteristics, to reduce the bandwidth to the image memory by 81.6%. In stereo matching vision task, this dissertation proposed the partial column reuse (PCR) and access reduction by expanding window (AREW) techniques, which were based on the access locality due to the algorithm’s flow, to significantly reduce bandwidth requirement for the proposed mini-census adaptive support weight (MCADSW) stereo matching architecture. The bandwidth utilization improving and bandwidth requirement reduction techniques studied in this dissertation can also be applied to other video coding or vision systems to improve system performance.