Software occlusion culling in Block Game

Eniko Fox implemented a CPU-based occlusion culling system for a voxel block game to help weak integrated GPUs run much faster. The system renders cube occluders into a low-resolution (256x128) depth buffer on a background thread, culling whole 16x16x16 chunks when their projected bounds are fully occluded. Chunks are preprocessed into a five-level mipmap of occluding subchunks (8x8x8 down to single blocks); opaque subchunks with visible faces become occluders gathered by distance rules and frustum checks. The pipeline is thread-safe by copying chunk positions and indices before the background pass, and conservative measures - shrinking occluders by one pixel and treating chunks with near-plane corners as visible - prevent false positives at low resolution.

Key implementation details drove the performance: occluders are rasterized by computing per-row min/max x extents from transformed cube corners and filling depth using the furthest corner depth, while occlusion candidates check against the nearest depth per pixel. Transform work was reduced by using a 3x3 view rotation plus additive translation and linear view-space depth instead of full 4x4 projection and 1/w divides, cutting arithmetic significantly. The result typically halves CPU rendering cost after frustum culling, removes 50-60% of chunks in open scenes and up to 95% in caves, and yields large frame-rate gains on low-end hardware, with a few edge cases where the cheap assumptions break down.