VRAM: compact the skinned vertex output (no UV copy, octahedral normal/tangent): 48 B -> 20 B

sky_line commented

2026-10-10 01:15:50 +00:00

Owner

The pre-skin writes 48 B per skinned vertex (DrawVertex: fp32 position 12, uv 8, normal 12, tangent 16). The rigged category was 534 MB on an RTX 3070 user's machine in a crowd; the skinned output is most of it.

Scope (no visible loss; positions stay exact fp32)

Stop copying UVs. Skinning never changes UVs: consumers read them from the source vertex (the 20 B QVertex the pre-skin already decodes them from), bit-identical to today. -8 B.
Octahedral normal + tangent at 16 bits per component (tangent sign kept): ~0.005 deg error, invisible in shading. 28 B -> 8 B.

Result: 48 B -> 20 B per skinned vertex (-58% of the skinned output). RT BLAS builds, motion vectors and rig weld are unaffected (positions unchanged). Every consumer of the skinned stream decodes the new layout: rigged raster (all passes), shadow / probe / mirror rigged draws, alpha sort, picking, RT copies, rig weld, motion vectors / history.

Position quantization (16-bit unorm per instance box) is a separate, later follow-up ticket.

The pre-skin writes 48 B per skinned vertex (DrawVertex: fp32 position 12, uv 8, normal 12, tangent 16). The rigged category was 534 MB on an RTX 3070 user's machine in a crowd; the skinned output is most of it. ## Scope (no visible loss; positions stay exact fp32) 1. **Stop copying UVs.** Skinning never changes UVs: consumers read them from the source vertex (the 20 B QVertex the pre-skin already decodes them from), bit-identical to today. -8 B. 2. **Octahedral normal + tangent at 16 bits per component** (tangent sign kept): ~0.005 deg error, invisible in shading. 28 B -> 8 B. Result: 48 B -> 20 B per skinned vertex (-58% of the skinned output). RT BLAS builds, motion vectors and rig weld are unaffected (positions unchanged). Every consumer of the skinned stream decodes the new layout: rigged raster (all passes), shadow / probe / mirror rigged draws, alpha sort, picking, RT copies, rig weld, motion vectors / history. Position quantization (16-bit unorm per instance box) is a separate, later follow-up ticket.

sky_line added the

labels

2026-10-10 01:15:50 +00:00

sky_line added this to the Slipstream project

2026-10-10 01:15:50 +00:00

sky_line changed title from ~~VRAM: quantize the skinned (pre-skin output) vertex buffers~~ to VRAM: compact the skinned vertex output (no UV copy, octahedral normal/tangent): 48 B -> 20 B

2026-10-10 01:20:14 +00:00

sky_line added

status

in-progress

and removed

triage

labels

2026-10-10 01:20:14 +00:00

sky_line referenced this issue

2026-10-10 01:20:14 +00:00

VRAM (later): quantize skinned positions (16-bit per-instance box) as a pressure-only ladder rung #617

sky_line referenced this issue from a commit

2026-10-10 03:31:53 +00:00

#616: compact skinned vertex 48 B -> 20 B (fp32 pos + octahedral normal / tangent)

sky_line referenced this issue from a commit

2026-10-10 03:31:53 +00:00

#616: clean_cull.comp RigCache comment wrap

sky_line referenced this issue

2026-10-10 04:53:21 +00:00

Explore: further texture memory reductions (RAM + VRAM) #263

sky_line commented

2026-10-10 05:33:00 +00:00

Author

Owner

Built and pushed (b401f5f8cb, 6713f114e1): skinned vertex output 48 B -> 20 B (fp32 position + octahedral normal/tangent). Shipped in 1.2.0. Only partially exercised in-world so far; moves to Testing until rigged avatars are checked for lighting/normal-map differences and the VRAM saving is measured in a crowd.

Built and pushed (b401f5f8cb, 6713f114e1): skinned vertex output 48 B -> 20 B (fp32 position + octahedral normal/tangent). Shipped in 1.2.0. Only partially exercised in-world so far; moves to Testing until rigged avatars are checked for lighting/normal-map differences and the VRAM saving is measured in a crowd.

sky_line added

and removed

labels

2026-10-10 05:33:00 +00:00

sky_line referenced this issue

2026-10-10 06:38:07 +00:00

Rig weld: also require similar skin weights before welding (don't weld cavities / onion-skin layers shut) #626

sky_line referenced this issue