<title>High-speed software raycasting on a dual-CPU PC using cache optimizations, MMX, streaming SIMD extensions, multithreading, and DirectX</title>

Abstract

We discuss work in progress towards a comprehensive software system for interactive volume rendering and the techniques we use to achieve high frame rates. The system was designed for Pentium-III CPUs and is currently running under Windows NT. It accepts three voxel formats and provides perspective views, on-the-fly classification and gradient shading. For high rendering speed, the entire recasting algorithm has been implemented and optimized in assembler, taking advantage of the MMX instructions and the Streaming SIMD Extensions of Pentium-III CPUs. These instructions are used to speed up frequently performed volume rendering operations such as tri-linear interpolations and ray-volume intersection test. The algorithm run is in parallel on two CPUs using multi-threading, and fast access to the frame buffer is established using DirectDraw. For further speed- up, early ray termination has been implemented as well. However, the major obstacle towards high performance is the limited memory bandwidth, even more so because volume rendering requires 3D access to the data set, and frequency access to tables. Thus, a memory layout was developed which maximizes the processor cache hit rate and keeps frequently used tables in the L1 cache, although at the cost of an increased memory consumption.