Compressed random-access trees for spatially coherent data

Abstract: Adaptive multiresolution hierarchies are highly efficient at representing spatially coherent graphics data. We introduce a framework for compressing such adaptive hierarchies using a compact randomly-accessible tree structure. Prior schemes have explored compressed trees, but nearly all involve entropy coding of a sequential traversal, thus preventing fine-grain random queries required by rendering algorithms. Instead, we use fixed-rate encoding for both the tree topology and its data. Key elements include the replacement of pointers by local offsets, a forested mipmap structure, vector quantization of inter-level residuals, and efficient coding of partially defined data. Both the offsets and codebook indices are stored as byte records for easy parsing by either CPU or GPU shaders. We show that continuous mipmapping over an adaptive tree is more efficient using primal subdivision than traditional dual subdivision. Finally, we demonstrate efficient compression of many data types including light maps, alpha mattes, distance fields, and HDR images.