CS448 Project Proposal: Texture Caches for Multitexturing

Motivation

In recent years, the use of multi-texturing as a cheap way to improve the realism of images in games has become widespread.  Initially, a technique called light-mapping was used to map both a texture and light map onto the same fragments.   The light map added lighting to the normally flat fragment to give the view an impression of specular lighting.  This technique required two textures, the light texture and the surface texture to be rasterized onto the same fragment.  Later, this was extended to bump mapping which requires multiple textures (a light vector, bump map in addition to the surface texture) to be textured onto the same fragments.  Today, Quake maintains at least 4 texturing levels -- a main texture, a light map, a fog map and either a detail map or a environment map.  Graphics cards are limited to 4 simultaneous textures.  We intend to explore the use of texture caches and the associated locality of multitextures to improve the performance in multi-pass textured images.

Bump Mapping:

Base Texture	Bump Map	Light Vector	Final Image

(From http://www.nvnews.net/previews/geforce3/multitexturing.shtml)

Goal of this Project

We indend to extend the work of Hakura and Gupta to explore the use of texture caches for improving the performance of multitextured images.  We hope to exploit the locality of images which are frequently rendered together and through placement reduce the amount of cache misses.  We also wish to explore the effects of higher degrees of multitexturing even though modern graphics cards are limited to 4 textures per fragment.

Key Technical Challenges and Approach

Our tasks are roughly outlined as follows:

• Leverage the simulation and data collection infra-structure of Hakura and Gupta.   This includes a simulator framework, library to trace OpenGL calls and a trace based cache simulator
• Find a set of images which exploit multitexturing and have interesting characteristics
• Conduct experiments using different texture layouts with different cache architectures
• Experiment with prefetching and/or compression
• Experiment with larger numbers of textures (higher degree of multitexturing resulting in possibly larger working sets)

In most cases, our measures of merit will be the number of cache misses and ultimately the resulting bandwidth.  We will begin by working with Hakura and looking for a good set of images to form a benchmark set.  We will then start with placement and cache architecture (associativity, cacheline size, etc) options and look at other methods of improving cache usage such as prefetching and compression.  If time allows, we will finish with exploring higher degrees of multitexturing.

Group Members

Francois Labonte:

• Find representative image set for benchmarks
• Work on OpenGL trace library
• Investigate the feasibility of compression

David Lie:

• Run pipeline and cache simulator
• Experiment with prefetching and texture placement
• Investiage higher order multitexturing