Hardware texture mapping is essential for real-time rendering. Unfortunately the memory bandwidth and latency often bounds performance in current graphics architectures. Bandwidth consumption can be reduced by compressing the texture map or by using a cache. However, the way a texture map occupies memory and how it is accessed affects the pattern of memory accesses, which in turn affects cache performance. Thus texture compression schemes and cache architectures must be designed in conjunction with each other. We define a sparse texture to be a texture where a substantial percentage of the texture is constant. Sparse textures are of interest as they occur often, and they are used as parts of more general texture compression schemes. We present a hardware compatible implementation of sparse textures based on B-tree indexing and explore cache designs for it. We demonstrate that it is possible to have the bandwidth consumption and miss rate due to the texture data alone scale with the area of the region of interest. We also show that the additional bandwidth consumption and hideable latency due to the B-tree indices are low. Furthermore, the caches necessary for these textures can be quite small.

Hardware texture mapping is essential for real-time rendering. Unfortunately the memory bandwidth and latency often bounds performance in current graphics architectures. Bandwidth consumption can be reduced by compressing the texture map or by using a cache. However, the way a texture map occupies memory and how it is accessed affects the pattern of memory accesses, which in turn affects cache performance. Thus texture compression schemes and cache architectures must be designed in conjunction with each other. We define a sparse texture to be a texture where a substantial percentage of the texture is constant. Sparse textures are of interest as they occur often, and they are used as parts of more general texture compression schemes. We present a hardware compatible implementation of sparse textures based on B-tree indexing and explore cache designs for it. We demonstrate that it is possible to have the bandwidth consumption and miss rate due to the texture data alone scale with the area of the region of interest. We also show that the additional bandwidth consumption and hideable latency due to the B-tree indices are low. Furthermore, the caches necessary for these textures can be quite small.