針對近似照明之高效率渲染技術

Abstract

照明對影像生成來說是個很重要的關鍵。從即時渲染，互動式渲染到光傳輸模擬，這些領域根據不同的應用目的試圖提升效率。在本論文中，我們將從硬體架構到演算法的角度介紹所提出的高效率渲染方法，主要為渲染高品質之照明效應。 本論文從一個感知功率的GPU繪圖管線架構開始介紹。從硬體的角度看來，我們想減少像素線程以節省順向渲染管線中之處理器功耗。主要因為照明通常是平順地分布在物體表面，所以我們想透過稀疏取樣像素執行像素照明，並且在繪圖管線中的光柵處理單元(ROP)以片狀單位(per tile basis)內插未被渲染之像素。所提出之屏幕近似照明方法實現在一顆十六核心行動裝置處理晶片原型，透過台積電45奈米製程實作。在我們的評估下，當執行順向照明時，處理器的執行比例可減少40%~50%。 基於硬體實作的方法是簡化的濾波技巧，並且受限於片狀單位處理。所以我們延伸自硬體方法，提出了一個基於屏幕空間的感知幾何濾波方法來重建稀疏照明之像素。透過隨機稀疏取樣，並照明屏幕空間上的像素，我們所提出之屏幕照明近似方法考慮幾何關係，譬如法向量以及位置做像素重建。實驗結果，平均來說，我們可以僅渲染20%~30%之屏幕像素並獲得令人滿意的重建品質。 更進一步，我們考慮了間接照明，以提供在即時或者互動渲染應用之全域照明效應來提升渲染真實度。除了直接照明，我們近似第一碰撞間接照明。所提出的技巧基於簡化後的點狀幾何，並透過體素格來索引。透過稀疏之光行進法，我們能檢查遮蔽體素，更進一步，我們提出了可視度濾波器來結合可視體素以平滑可視度估計。接著，我們能利用可能的點狀光源來近似第一碰撞間接照明。效能上來說，我們所提出之方法能提供間接照明之互動效能約為每秒1~2片幀。 最後，我們延伸點狀幾何來近似路徑追跡。為了有效率地找出光與幾何之交集點以連結路徑頂點，我們提出了一個混合查閱結構。這個查閱結構結合了點，叢集平面以及體素格。透過使用基於體素格的查閱，我們可以獲得一個粗略體素格，透過此粗略估計我們可以透過此查閱結構區域性地查閱並連結可能的路徑頂點。實驗結果，我們能在幾秒內快速減少生成影像之變異度，達到高品質的影像生成結果。 透過我們的實驗結果，從硬體到平行實作的角度，我們提出的方法在即時圖像到光傳輸模擬之領域都能達到很高的效能，同時兼顧並達到令人滿意之影像生成品質。總的來說，本論文可以視為主要在高效率渲染之近似照明領域的完整討論。

Illumination is essential for synthesizing images. From real-time rendering, interactive rendering to light transport simulation, these fields exploit efficiency in producing illumination according to different application targets. In this dissertation, we will introduce our efficient rendering techniques for rendering illumination with plausible quality, from architecture to algorithm perspective. The dissertation starts from exploiting a power-aware rendering architecture for a conventional GPU pipeline. From hardware perspective, we aim at reducing the pixel threads to save processor power dissipation in the forward rendering pipeline. Since illumination usually reveals smooth variation, instead of lighting every rasterized pixel, we sparsely sample pixels for lighting and interpolate the un-shaded pixels per tile basis during ROP phase. The screen-space approximated lighting technique is realized in a 16-core mobile GPU chip prototype, which is fabricated with TSMC 45nm technology. In our evaluation, the processor execution can reduce about 40% to 50% while performing forward lighting. The hardware-based approach is a simplified filtering technique and limited by the tile basis. Extended from the hardware-based approach, we propose a screen-space geometry-aware bilateral filter for sparsely lit pixels. With randomly illuminated screen pixels, the screen-space lighting approximation considers deferred geometric properties, including normal and position. Empirically, we can merely shade 20% to 30% screen pixels in average with satisfactory reconstruction quality. Moreover, indirect illumination is the soul of global illumination which enhances the visual realism for either real-time or interactive rendering applications. In addition to direct illumination, we approximate the first bounce indirect illumination. Our technique is based on a simplified point-based geometry, which is indexed via voxel grids. Through sparse ray marching for examining occluded voxels, we propose a visibility filter for associating batches of visible voxels to smooth visibility estimation. Therefore, the first bounce indirect illumination can be approximated with the possible point light sources. Regarding performance, the proposed technique can produce indirect illumination in an interactive rate from 1 to 2 fps. Ultimately, we extend the point-based geometry to approximate path tracing. To efficiently query the intersection for vertex connection, we propose a hybrid lookup hierarchy. The lookup structure is composed of point samples, clustered planes and voxel grids. Through utilizing a voxel-based query approach on coarse voxels, the rough estimate can be used for locally connecting the possible path vertex via the lookup structure. As a result, the variance of a synthesized image can be rapidly reduced in a few seconds with plausible visual quality. Our evaluation reveals compelling performance results, from architecture to parallel implementations, from real-time graphics to light transport simulation, meanwhile delivering plausible visual quality. In summary, the dissertation can be regarded as an extensive study in major efficient rendering topics for approximated illumination.