Two-phase occlusion culling: switch on/off automatically by how much it culls #806

New issue

Open

opened 2026-10-17 15:03:18 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-17 15:03:18 +00:00

Owner

Two-phase Hi-Z occlusion culling (RenderVKOcclusion2Phase) is now off by default (a911265b9b). In an open outdoor scene it costs more than it saves. On a GTX 1080 at 4K with Firestorm-equivalent settings, turning it off took the frame from 25.4 to 22.1 ms and from 23.6 to 20.9 ms in two A/B pairs, about 3 ms either way. The second pre-pass phase, the Hi-Z build and the occlusion test cost more than the little geometry they hide. In dense interiors and cities it should still pay off.

Goal: turn it on and off by itself, based on how much it actually culls in the current scene, without causing hitches or popping.

Things to get right

Signal. Measure the benefit from what the GPU already reports, a few frames late: instances or triangles rejected by the phase-2 test against those tested, and the GPU time of the pre-pass p2 + c_hiz + c_occl_p2 + c_occl_split markers. Compare against an estimate of the work it saves. Smooth the signal over seconds, not frames.
Hysteresis.
- Separate on and off thresholds, plus a minimum time in each state so camera turns can't make it flip-flop.
- While it is off, re-probe at a low rate: one measured frame now and then, or a cheap estimate, so it can find out when a scene would benefit.
- Re-evaluate right away on teleport or region change.
No hitches when switching. Enabling it must not create resources or rebuild descriptors in the frame it switches. Keep its buffers and Hi-Z allocated, or allocate them ahead of time while it is off, so a switch only changes which passes get recorded. Watch for the first-frame problem: the phase-1 "visible last frame" list is empty or stale right after it turns on.
Consistency. No popping. Phase 2 must still draw everything phase 1 missed, both on the switch frame and on the frame after.
Cluster LOD dependency. With two-phase and RenderVKOcclusionCull both off, no Hi-Z is built, so RenderVKClusterOcclusion is idle too. Decide whether a Hi-Z should still be built cheaply for cluster culling while two-phase is off, and measure that on its own.
Measure per GPU class. Pascal pays more for the extra passes than newer cards do, so the on/off thresholds may need to depend on the GPU. Test on the GTX 1080 box and on a current card, in an outdoor scene, a dense city and an interior.

Settings stay available for A/B: RenderVKOcclusion2Phase forces it on, and an auto mode would sit alongside it.

Two-phase Hi-Z occlusion culling (`RenderVKOcclusion2Phase`) is now **off by default** (a911265b9b). In an open outdoor scene it costs more than it saves. On a GTX 1080 at 4K with Firestorm-equivalent settings, turning it off took the frame from 25.4 to 22.1 ms and from 23.6 to 20.9 ms in two A/B pairs, about 3 ms either way. The second pre-pass phase, the Hi-Z build and the occlusion test cost more than the little geometry they hide. In dense interiors and cities it should still pay off. **Goal:** turn it on and off by itself, based on how much it actually culls in the current scene, without causing hitches or popping. **Things to get right** - **Signal.** Measure the benefit from what the GPU already reports, a few frames late: instances or triangles rejected by the phase-2 test against those tested, and the GPU time of the pre-pass `p2` + `c_hiz` + `c_occl_p2` + `c_occl_split` markers. Compare against an estimate of the work it saves. Smooth the signal over seconds, not frames. - **Hysteresis.** - Separate on and off thresholds, plus a minimum time in each state so camera turns can't make it flip-flop. - While it is off, re-probe at a low rate: one measured frame now and then, or a cheap estimate, so it can find out when a scene would benefit. - Re-evaluate right away on teleport or region change. - **No hitches when switching.** Enabling it must not create resources or rebuild descriptors in the frame it switches. Keep its buffers and Hi-Z allocated, or allocate them ahead of time while it is off, so a switch only changes which passes get recorded. Watch for the first-frame problem: the phase-1 "visible last frame" list is empty or stale right after it turns on. - **Consistency.** No popping. Phase 2 must still draw everything phase 1 missed, both on the switch frame and on the frame after. - **Cluster LOD dependency.** With two-phase and `RenderVKOcclusionCull` both off, no Hi-Z is built, so `RenderVKClusterOcclusion` is idle too. Decide whether a Hi-Z should still be built cheaply for cluster culling while two-phase is off, and measure that on its own. - **Measure per GPU class.** Pascal pays more for the extra passes than newer cards do, so the on/off thresholds may need to depend on the GPU. Test on the GTX 1080 box and on a current card, in an outdoor scene, a dense city and an interior. Settings stay available for A/B: `RenderVKOcclusion2Phase` forces it on, and an auto mode would sit alongside it.

sky_line added the

perf

rendering

labels

2026-10-17 15:03:18 +00:00

sky_line added this to the Slipstream project

2026-10-17 15:03:18 +00:00

sky_line added the

status

up-next

label

2026-10-17 18:10:11 +00:00