Cel P6: avatar mask — per-pixel avatar id from the GPU-driven pre-pass #549

New issue

Closed

opened 2026-10-08 06:28:04 +00:00 by sky_line · 1 comment

sky_line commented

2026-10-08 06:28:04 +00:00

Owner

Part of epic #542 (from #186). Design + code citations are in the epic; read it first.

Size: M (3-4 d)

Per-pixel avatar identity for cel shading and FX scope. GPU-driven only: no per-face CPU work.

Scope

R16_UINT id attachment (avatar slot + 1, 0 = none), written by the depth pre-pass twins (static + rigged, opaque + mask) when a consumer wants it. The pre-pass is forced on while needed (RenderVKDepthPrepass, :25051).
- Rigged = instance flag bit0, slot from SkinMeta.avatar_idx (llcleanskinstore.h:78-80).
- Worn rigid attachments = bit9, skin_id = joint-table slot (llcleaninstancestore.h:61-67, clean_draw.vert:118).
- Verify both share the avatar-slot index space (ensureAvatarSlot, llcleanskinstore.h:104-115).
- System avatar body included.
Alpha-blended rigged (pass 2: hair, sheer clothing): a small extra pass re-issuing pass 2's rigged indirect lists into the id target (depth-test against scene depth, no depth write, alpha ≥ threshold).
Per-avatar selection SSBO (bitset by avatar slot): self / friends / all / chosen. Updated on change only.
Display-res, unjittered resolve (same mapping as clean_depth_up.comp) for post consumers.
Edge cases: own avatar in mouselook (hidden), impostors / jelly dolls (none = no mask), HUD space (excluded), mirrors / probes (not needed).
Debug view: RenderVKDebugView value showing ids as colours.

Done when

Debug view shows exact, stable coverage for several avatars with mesh bodies, rigged + rigid attachments and alpha hair, under TAA, render scale < 1 and camera motion.
Present in snapshots (forward / offscreen path).
Cost ≤ 0.1 ms at 1440p with 20 avatars; 0 when no consumer.

Part of epic #542 (from #186). Design + code citations are in the epic; read it first. **Size:** M (3-4 d) Per-pixel avatar identity for cel shading and FX scope. GPU-driven only: no per-face CPU work. **Scope** - **R16_UINT id attachment** (`avatar slot + 1`, 0 = none), written by the depth pre-pass twins (static + rigged, opaque + mask) when a consumer wants it. The pre-pass is forced on while needed (`RenderVKDepthPrepass`, :25051). - Rigged = instance flag bit0, slot from `SkinMeta.avatar_idx` (`llcleanskinstore.h:78-80`). - Worn rigid attachments = bit9, `skin_id` = joint-table slot (`llcleaninstancestore.h:61-67`, `clean_draw.vert:118`). - **Verify both share the avatar-slot index space** (`ensureAvatarSlot`, `llcleanskinstore.h:104-115`). - System avatar body included. - **Alpha-blended rigged** (pass 2: hair, sheer clothing): a small extra pass re-issuing pass 2's rigged indirect lists into the id target (depth-test against scene depth, no depth write, alpha ≥ threshold). - **Per-avatar selection SSBO** (bitset by avatar slot): self / friends / all / chosen. Updated on change only. - **Display-res, unjittered resolve** (same mapping as `clean_depth_up.comp`) for post consumers. - **Edge cases:** own avatar in mouselook (hidden), impostors / jelly dolls (none = no mask), HUD space (excluded), mirrors / probes (not needed). - **Debug view:** `RenderVKDebugView` value showing ids as colours. **Done when** - [ ] Debug view shows exact, stable coverage for several avatars with mesh bodies, rigged + rigid attachments and alpha hair, under TAA, render scale < 1 and camera motion. - [ ] Present in snapshots (forward / offscreen path). - [ ] Cost ≤ 0.1 ms at 1440p with 20 avatars; 0 when no consumer.

sky_line added the

labels

2026-10-08 06:28:04 +00:00

sky_line added this to the Slipstream project

2026-10-08 06:28:04 +00:00

sky_line referenced this issue

2026-10-08 06:28:04 +00:00

Epic: built-in effects stack (ReShade FX compatible) + cel-shading avatars #542