X Accelerating Virtual Texturing Using CUDA

For a long time games have used textures to add surface details and diversity to their virtual worlds. These textures usually consist of a basic set of different surface appearances which are then composed at runtime by shaders to calculate the final surface appearance. Recently there has been an increased interest in virtual texturing technologies [Barrett 08, Lefebre et al. 04,Mittring 08]. Virtual texturing allows very large textures (in the order of one gigapixel) to be applied to the game’s geometry while still remaining within the limits of today’s hardware. This allows far more varied worlds than can be achieved with composing tiling textures at a lower or comparable render cost. One of the first commercial systems to employ these techniques was id Soft-ware’s Mega Texturing technology [van Waveren 08] as used in the game Enemy Territory Quake Wars. This technology is loosely based on clipmapping [Tanner et al. 98] and shares the same limitation that only a single rectangle in texture space is available at the highest resolution. Hence its use is limited to nearly planar geometries with regular texture coordinates. In their upcoming game Rage, these limitations have been addressed by adopting a fully functional virtual texturing system [van Waveren 09]. Extending this technique to arbitrary geometries and texture coordinate mappings adds a substantial overhead to the technique. Besides the streaming, (de)compression and texture updates, the additional cost of determining what parts of the texture are referenced by the rendered frame is introduced. Exist-ing systems usually involve a lot of expensive CPU work and CPU to GPU data transfers