Skip to content

docs: re-profile ALab and the Moana island at 0.5.0 - #211

Merged
doubleailes merged 3 commits into
mainfrom
docs/production-profiles-0.5.0
Oct 5, 2026
Merged

doubleailes merged 3 commits into
mainfrom
docs/production-profiles-0.5.0

Conversation

@doubleailes

Copy link
Copy Markdown
Owner

Summary

Production-scene measurements refreshed at 4143c58 (0.5.0), on the 72-vCPU machine, 2026-10-05. Docs only.

ALab frame 1004 (docs/alab_profile.md)

  • Render 3:20.7 → 23.4 s; the import is now 88% of the frame. Peak 33.2 → 28.7 GiB.
  • Texture lookups are 43% of render thread time, down from 89%, at 1.57 µs per eval. Shading / Raytrace / Integrator is 48 / 40 / 11.
  • 72 threads render 5.4× faster than 8, against 1.25×.
  • Light and shadow links are now read (3 occluder classes). The four lgt_screenLights are dropped because their lightLink names no receiver in the composed stage. Occluded shadow rays: 96% → 82%.

Moana island (docs/moana_profile.md)

  • The full default render (128 spp) takes 29.3 s, where the first profile took 5:24 at 4 spp. Peak 51.5 → 31.6 GiB, with displacement read.
  • Still 90% traversal, at 17.7 µs per closest hit; the import is 91% of the frame.
  • The distantPalm_key light links resolve as intended.
  • 35 archive prototypes contribute no geometry. This is unchanged since 2026-10-02 and not investigated.

Ptex streaming on the island (docs/ptex_streaming.md)

  • MIPSPACE=file streaming against preload, interleaved: resident 7.34 → 0.61 GiB, Load assets 1:33 → 28 s, peak 31.6 → 23.9 GiB.
  • Render +1.8% / +3.2%, within noise; the whole run is 67 s shorter.

The first profiles are kept as dated history sections. The matching figures in openspec/specs/usd-scene-import/design.md and in the CRUST_PTEX_STREAM_MIPSPACE user docs are updated.

Not updated:

  • The Moana sky-rig colour table is still marked stale.
  • docs/light_sampling.md §3.12 has ALab's learned-vs-power comparison. It predates shadow links and probably needs re-measuring.
  • openspec/changes/stream-ptex-by-default/proposal.md still quotes the old +1.2%.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FXGCHSpPr1joewseS5422u

doubleailes and others added 3 commits October 5, 2026 11:01
Same recipe as the first profile (--stats, --profile, and the
72/8-thread pair), measured 2026-10-05 at 4143c58. The render fell from
3:20.7 to 23.4 s and is now 12% of the frame: USD composition is the
other 88%. Texture lookups are still the largest render section (43%,
1.57 us per eval) but no longer swamp it; Shading/Raytrace/Integrator is
48/40/11. 72 threads now render 5.4x faster than 8, against 1.25x.

Record that light and shadow links are read (3 occluder classes), that
the four lgt_screenLights are dropped because their lightLink names no
receiver in the composed stage, and the NEE, path and streaming counters.
The first profile and the fixes it led to are kept as a history section.

Bring the ALab figures in the usd-scene-import design record up to date.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FXGCHSpPr1joewseS5422u
Measured 2026-10-05 at 4143c58, at the true defaults (128 spp,
displacement on, Ptex preloaded): the render takes 29.3 s of a 5:42 run
and peaks at 31.6 GiB, against 5:24 at 4 spp and 51.5 GiB in the first
profile. It is still 90% traversal, at 17.7 us per closest hit; the
import is 91% of the frame.

Record that the island's light links resolve as the rig intends (the
distantPalm_key lights now light 1 of 2 receiver classes), that 35
archive prototypes contribute no geometry (unchanged since 2026-10-02,
not investigated), and the ray, path and adaptive counters. The first
profile is kept as a history section; the later dated sections stay.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FXGCHSpPr1joewseS5422u
The island table was taken before the traversal fix, on a 14-minute
render. Measured 2026-10-05 at 4143c58, 8 spp, two runs per side
alternating, CRUST_PTEX_STREAM=1 CRUST_PTEX_STREAM_MIPSPACE=file against
preloading: Ptex resident 7.34 -> 0.61 GiB, Load assets 1:33 -> 28 s,
peak 31.6 -> 23.9 GiB, Render +1.8% / +3.2% (within noise), the whole
run 67 s shorter. Peak now falls by the full traverse saving, since the
commit's transient is small. The first measurement is kept in one line.

Update the same figures in the usd-scene-import design record and the
CRUST_PTEX_STREAM_MIPSPACE user documentation.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FXGCHSpPr1joewseS5422u
@qodo-code-review

Copy link
Copy Markdown
Contributor

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (0) 📎 Requirement gaps (0)

Grey Divider

Great, no issues found!

Qodo reviewed your code and found no material issues that require review

Grey Divider

Tip of the day
💡 Did you know, you can turn on the rule miner and Qodo learns your standards from review history

More tips ↗ | Customize Qodo ↗ | Qodo docs ↗

Grey Divider

Qodo Logo

@doubleailes
doubleailes merged commit 6a08ab1 into main Oct 5, 2026
12 checks passed
@doubleailes
doubleailes deleted the docs/production-profiles-0.5.0 branch October 5, 2026 09:49
@qodo-code-review

Copy link
Copy Markdown
Contributor

PR Summary by Qodo

Re-profile ALab and Moana production scenes at 0.5.0

📝 Documentation 🕐 40+ Minutes

Grey Divider

AI Description

• Refresh ALab and Moana profiles to show current render, import, memory, and lighting behavior.
• Re-measure Moana Ptex streaming, including its memory savings and mip-level colour trade-off.
• Preserve earlier measurements as dated history and align the import design and user reference.
Diagram

graph TD
  ALab["ALab profile"] --> Design["Import design"] --> Reference["Environment reference"]
  Moana["Moana profile"] --> Ptex["Ptex benchmark"] --> Design
  Ptex --> Reference
Loading
High-Level Assessment

Keep the detailed, dated measurements in the existing profiles and update the shorter design and user-facing summaries. A single central table could reduce repeated figures, but would make these established references less useful in place.

Files changed (5) +457 / -108

Documentation (5) +457 / -108
alab_profile.mdAdd the 0.5.0 ALab frame profile +199/-22

Add the 0.5.0 ALab frame profile

• Documents current frame timings, memory, render sections, thread scaling, and light and shadow links. Moves the first profile and its resulting fixes into a dated history section.

docs/alab_profile.md

moana_profile.mdAdd the 0.5.0 Moana island profile +178/-26

Add the 0.5.0 Moana island profile

• Records a full-default 128-spp run, including import and traversal costs, memory, displacement, light links, and unresolved geometry-free prototypes. Retains the original profile and subsequent measurements as dated history.

docs/moana_profile.md

ptex_streaming.mdRefresh the island Ptex streaming comparison +51/-36

Refresh the island Ptex streaming comparison

• Replaces the primary preload-versus-streaming figures with interleaved 0.5.0 measurements and retains the earlier result for comparison. Explains the reduced load time and memory, near-noise render cost, and stored-mip colour bias.

docs/ptex_streaming.md

design.mdAlign import design figures with the new profiles +28/-23

Align import design figures with the new profiles

• Updates the ALab, Moana, and Ptex streaming summaries with current timings, memory, and bottlenecks. Keeps earlier benchmark figures identified as historical measurements.

openspec/specs/usd-scene-import/design.md

environment-variables.mdUpdate Ptex file-mipspace memory figures +1/-1

Update Ptex file-mipspace memory figures

• Refreshes the Moana residency and peak-memory savings quoted for 'CRUST_PTEX_STREAM_MIPSPACE=file', while retaining the warning about its colour bias.

site/content/docs/reference/environment-variables.md

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant