Right-size the avatar joint-slot tables (no fixed cap) #762

New issue

Open

opened 2026-10-14 17:11:32 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-14 17:11:32 +00:00

Owner

Today

Joint table + per-slot CPU arrays (LLCleanSkinStore::MAX_AVATARS), ~10-14 KB per slot.
Avatar visibility table (kAvVisSlots) with compile-time section offsets (kCasterBase, kAvBoxBase, kAvLodBase ...) shared by C++ and GLSL.
Cluster draw-list avatar masks (rig_bits, fixed words in a uniform struct).
Cluster instance flags packing the slot index into a fixed bit field (CLI1B_AV_SHIFT).
RT per-avatar masks, admitted / layer bitsets, kSkinAvRows / kRigAvRows.
~170 C++ references in 15 files, ~50 in 36 shaders.

Goal

Start small (e.g. 64 slots), grow by doubling with no fixed ceiling short of the VRAM budget; shrink after a grace period like the other stores (free on liveness).
Runtime slot capacity read by shaders (no compile-time 256/1024); masks as word buffers indexed by slot >> 5; slot index in its own word (or a field wide enough for any capacity).
Growth = reallocate + copy + descriptor update (same pattern as the rigged skin buffers); at most a one-frame blip, only past the previous high-water.
Log live / capacity / peak slots.
Debug check that clamps + logs any out-of-range slot index (a missed reader would otherwise index past a buffer on the GPU).
Allocation failure never crashes (retry later, then drop + log).

Notes

Past the cap today, with cluster LOD off a new character falls back to static (unskinned) geometry; with cluster LOD on it is cluster-owned with no slot and does not draw. Right-sizing removes the cap rather than adding a fallback path.

The renderer caches skinning joints for animated characters (avatars, animesh control avatars, and flexi prims packed 24 chains per slot) in a fixed table of avatar slots. The cap was a hard-coded 256 (raised to 1024 as a stopgap). Replace the fixed cap with buffers that are sized to demand and grow. ## Today - Joint table + per-slot CPU arrays (`LLCleanSkinStore::MAX_AVATARS`), ~10-14 KB per slot. - Avatar visibility table (`kAvVisSlots`) with compile-time section offsets (`kCasterBase`, `kAvBoxBase`, `kAvLodBase` ...) shared by C++ and GLSL. - Cluster draw-list avatar masks (`rig_bits`, fixed words in a uniform struct). - Cluster instance flags packing the slot index into a fixed bit field (`CLI1B_AV_SHIFT`). - RT per-avatar masks, admitted / layer bitsets, `kSkinAvRows` / `kRigAvRows`. - ~170 C++ references in 15 files, ~50 in 36 shaders. ## Goal - Start small (e.g. 64 slots), grow by doubling with no fixed ceiling short of the VRAM budget; shrink after a grace period like the other stores (free on liveness). - Runtime slot capacity read by shaders (no compile-time 256/1024); masks as word buffers indexed by `slot >> 5`; slot index in its own word (or a field wide enough for any capacity). - Growth = reallocate + copy + descriptor update (same pattern as the rigged skin buffers); at most a one-frame blip, only past the previous high-water. - Log live / capacity / peak slots. - Debug check that clamps + logs any out-of-range slot index (a missed reader would otherwise index past a buffer on the GPU). - Allocation failure never crashes (retry later, then drop + log). ## Notes - Past the cap today, with cluster LOD off a new character falls back to static (unskinned) geometry; with cluster LOD on it is cluster-owned with no slot and does not draw. Right-sizing removes the cap rather than adding a fallback path.

sky_line added the

perf

avatars

rendering

labels

2026-10-14 17:11:32 +00:00

sky_line added this to the Slipstream project

2026-10-14 17:11:32 +00:00

sky_line referenced this issue from a commit

2026-10-15 03:16:15 +00:00

Avatar slot cap 256 -> 1024 (joint table, av_vis, cluster LOD, RT, weld, cel)