可動態重組之材質處理單元

Abstract

材質過濾單元在3-D繪圖處理器中佔有相當比重的成本，而且其所佔比例因為以下兩個原因持續增加中。從Shader Model 3.0 及High Dynamic Range的規格開始支援浮點格式的材質過濾運算，且3-D繪圖處理器架構中的材質過濾單元數量(即材質處理單元數量，一個材質處理單元包含了一個材質過濾單元和其他元件)也持續增加。故在擬真畫面的即時顯像下的低成本之材質過濾單元是迫切需要的。材質過濾類別的需求在執行期間會有所改變，且各種材質過濾類別彼此間具有大量共用的運算，故一個可以彈性地在各種材質過濾類別做功能切換的材質過濾單元可以大量節省其面積。動態可重組式運算和ASIC相比，提供了一個在較低成本下能得到較高運算速度的機會。此外，需要相同材質過濾類別需求的連續次數很長也意味著重組代價相對也少。在本篇論文中，我們提出了一個具低面積速度乘積的可動態重組之材質過濾單元設計。首先，我們針對在單一材質處理單元內的材質過濾單元提出一個類似divide-and-conquer的低面積速度乘積設計。接著，我們整合多個材質處理單元內的上述材質過濾單元，以組成一個較大並且跨多個材質處理單元的材質過濾單元，並且設計一套類以round-robin的方法提升整合過後的材質處理單元之使用率。實驗結果顯示此較大並且跨多個材質處理單元的材質過濾單元相較於多個單一材質處理單元內的材質過濾單元，在僅增加了2.5%的執行時間下,節省了14.2%的面積以及12.1%的面積速度乘積。

Texture filters are one of cost intensive parts in 3-D Graphics Processing Unit architecture and cost of them is increasing due to revealed floating point filtering techniques in Microsoft Shader Model 3.0 with High Dynamic Range feature and increasing number of texture filters (i.e. # of texture unit, a texture unit is composed of a texture filter and other components) required in GPU architecture. Therefore, a low-cost texture filter design in realistic and real-time rendering is required. The usage of filter types varies over run time. Furthermore, there are a large number of shared operations between each filter type. An all-purpose texture filter which can switch flexibly for all filter type can save significant area. Run-time reconfigurable computing offers an opportunity of obtaining high computing speed at lower hardware costs, compared to ASIC. Long run-length of data requiring the same filter type implies small reconfigurable overhead. In this thesis, a run-time reconfigurable texture filter design with low area-time product is proposed. First, a divide-and-conquer-like design for a low area-time product texture filter within a texture unit is proposed. We form a larger texture filter cross multiple texture units to improve its area-time product by integrating the texture filters within multiple texture units, which is the second design. A round-robin-like design to increase the utilization of a reconfigurable texture unit is proposed. The experimental result shows the texture filter cross multiple texture units design achieve 14.2% area saving with 2.5% longer execution time and 12.1% area-time product reduction compared to texture filters within multiple texture units design.