[rendering] Rigged BLAS refit: importance-based cadence + skip refit on near-unchanged poses #1007

New issue

Open

opened 2026-10-24 01:23:22 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-24 01:23:22 +00:00

Owner

Busy party (RTX 5090, 65 avatars traced): q_rt_build ~3.7 ms/frame on the async queue ahead of the lighting chain; c_blas_refit 2.04 ms = 115 of 130 rigged groups refit every frame (10k faces, 831 MB rigged BLAS); c_preskin 0.91 (shared with raster), c_tlas 0.26. Existing gating already exists ("gated 13 (held + moving 0)", "no-self-shadow (throttled) 27") but almost everything refits every frame.

1. Refit cadence by importance

Every frame: avatars within ~20 m of the camera, or on screen and large (projected size threshold), or casting visible RT shadows near the camera, plus your own avatar.
Mid range every 2nd frame, far every 4th; staggered (group id / phase) so the per-frame cost stays flat, not spiky.
GI and RT reflections tolerate 1-3 frames of staleness; distant avatars' RT shadows lagging ~50 ms is not visible.
Teleport / region change / an avatar appearing or its geometry changing: refit (or build) at once.
The TLAS instance transform still updates every frame (only the BLAS contents lag); make sure a skipped refit never leaves the BLAS reading skinned positions that were overwritten in place for raster (double-buffer or snapshot as the existing rigPos design requires).

2. Skip refit when the pose barely changed

Compare the skin palette (bone matrices) against the palette used at the group's last refit; under a threshold (max bone translation delta in metres / rotation in degrees, scaled by distance) skip the refit this frame. Idle / standing loops benefit; dancing avatars mostly won't.
Hard cap on consecutive skips (staleness bound).

Settings (live): cadence on/off (A/B = every frame), distance bands, pose threshold, max staleness. Stats: refits/frame, skipped by cadence, skipped by pose, max staleness. Expected: ~2.0 -> 0.7-0.9 ms in this scene.

Busy party (RTX 5090, 65 avatars traced): `q_rt_build` ~3.7 ms/frame on the async queue ahead of the lighting chain; `c_blas_refit` **2.04 ms** = 115 of 130 rigged groups refit every frame (10k faces, 831 MB rigged BLAS); `c_preskin` 0.91 (shared with raster), `c_tlas` 0.26. Existing gating already exists ("gated 13 (held + moving 0)", "no-self-shadow (throttled) 27") but almost everything refits every frame. **1. Refit cadence by importance** - Every frame: avatars within ~20 m of the camera, or on screen and large (projected size threshold), or casting visible RT shadows near the camera, plus your own avatar. - Mid range every 2nd frame, far every 4th; **staggered** (group id / phase) so the per-frame cost stays flat, not spiky. - GI and RT reflections tolerate 1-3 frames of staleness; distant avatars' RT shadows lagging ~50 ms is not visible. - Teleport / region change / an avatar appearing or its geometry changing: refit (or build) at once. - The TLAS instance transform still updates every frame (only the BLAS contents lag); make sure a skipped refit never leaves the BLAS reading skinned positions that were overwritten in place for raster (double-buffer or snapshot as the existing rigPos design requires). **2. Skip refit when the pose barely changed** - Compare the skin palette (bone matrices) against the palette used at the group's last refit; under a threshold (max bone translation delta in metres / rotation in degrees, scaled by distance) skip the refit this frame. Idle / standing loops benefit; dancing avatars mostly won't. - Hard cap on consecutive skips (staleness bound). Settings (live): cadence on/off (A/B = every frame), distance bands, pose threshold, max staleness. Stats: refits/frame, skipped by cadence, skipped by pose, max staleness. Expected: ~2.0 -> 0.7-0.9 ms in this scene.

sky_line added the

labels

2026-10-24 01:23:22 +00:00

sky_line added this to the Slipstream project

2026-10-24 01:23:22 +00:00

sky_line referenced this issue

2026-10-24 01:31:07 +00:00

[rendering] Cluster LOD perf: close the gap to legacy in avatar-heavy scenes (RT cut, shadow rig cull, pre-cluster pass, raster) #1008