Performance Improvements for the Rasterization Pipeline

Jon Hasselgren
Lund University


Performance improvements are always needed in computer graphics. Better performance frees up computational resources which can be used to increase the level of realism in the rendered images. Despite a very rapid development of graphics hardware, we are still far from the point where photo-realistic rendering can be done in real time. Therefore, better algorithms must be developed in order to advance the field.

In this thesis, we present several new algorithms targeted for the rasterization pipeline, which is, at the time of writing, the de-facto standard rendering algorithm used in graphics hardware. We focus on three areas in the pipeline. The first is the rasterization step where we present efficient sampling strategies which can improve performance, or enable new algorithms to be implemented. This includes an investigation of very inexpensive, but relatively high quality, sampling schemes, an algorithm for conservative rasterization through area sampling, and an algorithm for efficiently rasterizing multiple views which can be useful for holographic displays, and other multi-view displays.

The second area of focus is on compression algorithms. Compression is frequently used in graphics hardware in order to minimize memory traffic and thereby increase performance. We present efficient algorithms for high dynamic range texture compression, depth buffer compression, and color buffer compression. The inner workings of buffer compression have been kept secret by hardware vendors, and the only available information is available through patents. Therefore, we also present surveys of the prior art in buffer compression.

The final area we focus on is programmable culling. Culling is an old concept in computer graphics, and it is often necessary to include some form of culling to make an algorithm truly efficient. However, current graphics hardware only has a few fixed function culling algorithms implemented. This often makes it impossible, or impractical, to implement culling in new rendering algorithms that rely on graphics hardware. We present a methodology for automatically deriving culling algorithms from shader programs, and we show how this can be used to perform culling both on a pixel level, and on a geometry or vertex level. In both cases, we show that our automatic culling can greatly improve rendering performance for a wide variety of scenes.


[ Thesis ]