g023

40,000 small objects in a city of grey blocks. Left draws everything inside the view: 12,364. Right tests chunks, then objects, against a Hi-Z depth pyramid: 1,953. Watch the split 👇 ▸ The problem Frustum culling keeps everything inside the view cone, including everything behind the next wall. Walking down a street, most of that is invisible, but it is still sent to the GPU and still costs vertices. ▸ How it works The source (a Unity URP renderer feature) does it in three steps: 1. Copy the depth buffer into a 512 texture and build mips. Each texel of the next mip keeps the farthest of the 2x2 below it, so one texel answers "what is the farthest thing covering this area?" 2. A compute shader projects every instance, picks a mip, and reads the Hi-Z there. If the instance is farther than the farthest occluder, it is hidden. Off-screen instances go too. 3. Survivors are appended to a buffer and drawn with one indirect instanced draw. I added a chunk level: the map is cut into 1,024 cells of 10 m. Chunk boxes are tested first, and objects in a rejected chunk are never tested at all. ▸ What I hit WebGL2 has no compute shaders and no indirect draw. My tests run in fragment passes, one pixel per chunk or object. I read the results back and pack the survivors into instanced meshes on the CPU. The source tests one centre point per instance with a distance-based mip. A centre point can hide an object that still peeks out, so I test the whole bounding box and pick the mip from its screen size. My Hi-Z is 512 x 256 from a 960 x 1080 view. The first level takes the farthest depth over each texel's whole footprint, so it never claims more cover than the full-res depth shows. ▸ My build three.js (r187), WebGL2. 254 buildings are the only occluders in the Hi-Z. 40,000 objects of 3 kinds: 80, 32 and 14 triangles. At 3 s: left draws 12,364 objects, 600,872 triangles. Right keeps 132 of 1,024 chunks (216 rejected by Hi-Z), tests 5,474 objects and draws 1,953, 93,480 triangles. ▸ Trade-offs + Big cuts wherever walls hide most of the view + The chunk test skips most per-object tests - Occluders must be in depth before the test - Reading results back stalls the GPU; the source avoids that with indirect draws - Small objects never hide anything - I rendered on a software GPU, so counts only, no timings ▸ When to use it Cities, interiors and dense props behind big occluders. In open fields frustum culling already does most of the work.