Explore: further texture memory reductions (RAM + VRAM)

sky_line commented

2026-10-03 04:18:36 +00:00

Owner

Current numbers (same session):

VRAM: textures 7.7 GB of a 15.4 GB process total. Budget 14.2 GB, bias 0; 1,638 textures at full res, 7,200 images; all-full would be 33 GB.
RAM: 2.7 GB of CPU texture copies (9,069). The earlier baseline (#159-#164) had 16.7 GB of CPU copies, so most of that is already fixed.

Ideas to evaluate:

GPU block compression (BC7 for colour, BC5 for normals, BC4 for masks) transcoded on upload, possibly in compute: 4:1 vs RGBA8, with a large VRAM and bandwidth win. Needs a quality check and transcode cost measurement (CPU vs GPU).
Screen-space-driven residency: choose each texture's mip from measured on-screen texel density (GPU feedback, e.g. the tex-demand readback) rather than LL's pixel-area heuristics, so textures that never reach full-res density never get it.
Drop CPU copies entirely after upload where nothing needs readback (keep only for sculpts / alpha-mask analysis / picking).
Deduplicate identical-content textures (same pixels, different UUID); avatar bakes/skins are common duplicates.
Normal / specular map resolution caps relative to diffuse.
Sparse/partially-resident textures for very large (2K/4K) images.
Tighter residency on 8 GB cards (see #164).

Deliverable: measurements per idea (VRAM saved, quality impact, CPU/GPU cost) on a heavy scene, then a ranked plan.

Current numbers (same session): - VRAM: textures 7.7 GB of a 15.4 GB process total. Budget 14.2 GB, bias 0; 1,638 textures at full res, 7,200 images; all-full would be 33 GB. - RAM: 2.7 GB of CPU texture copies (9,069). The earlier baseline (#159-#164) had 16.7 GB of CPU copies, so most of that is already fixed. Ideas to evaluate: - **GPU block compression** (BC7 for colour, BC5 for normals, BC4 for masks) transcoded on upload, possibly in compute: 4:1 vs RGBA8, with a large VRAM and bandwidth win. Needs a quality check and transcode cost measurement (CPU vs GPU). - **Screen-space-driven residency:** choose each texture's mip from measured on-screen texel density (GPU feedback, e.g. the tex-demand readback) rather than LL's pixel-area heuristics, so textures that never reach full-res density never get it. - **Drop CPU copies entirely** after upload where nothing needs readback (keep only for sculpts / alpha-mask analysis / picking). - **Deduplicate** identical-content textures (same pixels, different UUID); avatar bakes/skins are common duplicates. - **Normal / specular map resolution caps** relative to diffuse. - **Sparse/partially-resident textures** for very large (2K/4K) images. - **Tighter residency on 8 GB cards** (see #164). Deliverable: measurements per idea (VRAM saved, quality impact, CPU/GPU cost) on a heavy scene, then a ranked plan.

sky_line added the

perf

rendering

labels

2026-10-03 04:18:36 +00:00

sky_line referenced this issue

2026-10-03 23:47:08 +00:00

GPU-ready texture cache: store block-compressed textures on disk (skip J2C decode + encode on reload) #274