Abstract
Traditionally, sort-middle is a technique that has been difficult to attain on clusters because of the tight coupling of geometry and rasterization processes on commodity graphics hardware. In this paper, we describe the implementation of a new sort-middle approach for performing immediate-mode rendering in Chromium. The Chromium Rendering System is used extensively to drive multi-projector displays on PC clusters with inexpensive commodity graphics components. By default, Chromium uses a sort-first approach to distribute rendering work to individual nodes in a PC cluster. While this sort-first approach works effectively in retainedmode rendering, it suffers from various network bottlenecks when rendering in immediate-mode. Current techniques avoid these bottlenecks by sorting vertex data as a pre-processing step and grouping vertices into specific bounding boxes, using Chromium's bounding box extension. These steps may be expensive, especially if the dataset is dynamic. In our approach, we utilize standard programmable graphics hardware and extend standard APIs to achieve a separation in the rendering pipeline. The pre-processing of vertex data or the grouping of vertices into bounding boxes are not required. Additionally, the amount of OpenGL state commands transmitted through the network are reduced. Our results indicate that the approach can attain twice the frame rates as compared to Chromium's sort-first approach when rendering in immediatemode.