Skip padding pixels during GSplat sorting and rendering

[mvaligursky]

Summary

• CPU sort worker now operates only on active splats, completely skipping work-buffer padding pixels. An indexMap (built once per world-state change, not per-sort) maps compact splat indices back to work-buffer pixel positions, so the sorted order buffer and distance arrays are sized to totalActiveSplats instead of totalUsedPixels.
• GPU compaction flag pass rejects padding pixels via a sentinel pcNodeIndex value (0xFFFFFFFF), preventing stale data in unused work-buffer regions from being treated as visible splats.
• Work-buffer copy shaders (GLSL + WGSL) now write sentinel values for pcId and pcNodeIndex in the padding branch, ensuring padding pixels are never mistaken for real splats regardless of sort path.
• Last partial-row sub-draw is extended to full row width so the fragment shader covers trailing padding pixels, preventing stale data from previous LODs causing ghost splats with GPU sorting.
• GSplat count stat now reports only active splats, excluding padding.

Technical Details

• GSplatWorldState tracks totalActiveSplats alongside totalUsedPixels
• GSplatManager passes totalActiveSplats for CPU sort paths and GPU compaction element counts
• Sort worker builds indexMap in the intervals handler and uses it during counting sort output
• GSplatInfo.updateSubDraws() extends colEnd of the last partial row to textureWidth
• compute-gsplat-compact-flag.js checks for sentinel nodeIdx == 0xFFFFFFFF before visibility lookup

[Copilot]

Pull request overview

This PR updates the unified GSplat pipeline to consistently treat work-buffer padding pixels as non-splats, reducing CPU sorting work and preventing GPU-side “ghost splats” caused by stale padding data.

Changes:

• Introduces totalActiveSplats (excluding per-placement row padding) and routes CPU sorting/CPU-compaction to operate only on active splats.
• Adds sentinel pcNodeIndex = 0xFFFFFFFF (and related shader writes) so GPU compaction/visibility rejects padding pixels reliably.
• Adjusts work-buffer sub-draw generation to cover trailing padding pixels so they get overwritten instead of retaining stale data.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/scene/shader-lib/wgsl/chunks/gsplat/frag/gsplatCopyToWorkbuffer.js	Writes pcId and a sentinel pcNodeIndex for out-of-bounds (padding) pixels in WGSL.
src/scene/shader-lib/wgsl/chunks/gsplat/compute-gsplat-compact-flag.js	Treats sentinel pcNodeIndex pixels as always invisible during GPU compaction flagging.
src/scene/shader-lib/glsl/chunks/gsplat/frag/gsplatCopyToWorkbuffer.js	Mirrors padding sentinel writes for GLSL.
src/scene/gsplat-unified/gsplat-world-state.js	Tracks totalActiveSplats in addition to totalUsedPixels.
src/scene/gsplat-unified/gsplat-unified-sort-worker.js	CPU worker sorts only active splats and maps compact indices back to work-buffer pixels via indexMap.
src/scene/gsplat-unified/gsplat-manager.js	Threads totalActiveSplats through CPU sort/compaction paths and updates stats to exclude padding.
src/scene/gsplat-unified/gsplat-info.js	Extends a sub-draw region to ensure trailing padding pixels are covered by fragment writes.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.