erik/acdream
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

fix #225: stabilize render pacing and frame CPU

Browse source 
Replace scheduler-quantized software sleeps with a reusable Windows high-resolution deadline timer, expose pacing in the frame profiler, and make shutdown wake every persistent mesh worker without losing the shared signal.

Preserve retail alpha order while using a stable radix, skip duplicate deferred-alpha SSBO packing, pack light sets, cache static selection descriptors, and retire historical material groups at the whole-frame boundary. The fixed dense-Caul sample improved from roughly 9-12 ms CPU to 5.3-6.2 ms without reducing visual quality.

Release build succeeds with zero warnings and all 6,300 tests pass with five intentional skips. Three independent retail, architecture, and adversarial reviews are clean; the post-review connected route remains pending because local ACE is offline.

Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Erik 2026-07-19 06:29:30 +02:00
parent 47d7086a74
commit 3718e341be
17 changed files with 1103 additions and 225 deletions
Download patch file Download diff file Expand all files Collapse all files

26
docs/ISSUES.md
View file

@ -117,6 +117,28 @@ balanced at 1,715/1,715 and composite physical residency remained below its
128 MiB ceiling. The older 32 FPS comparison run was RDP-refresh-capped and is
used only for memory comparison. The lifestone/particle visual gate remains.
The 141.8 FPS result above is one destination-specific dwell, not a universal
throughput claim. A fixed dense-Caul capture with about 21,024 visible entities
later isolated a render-thread bottleneck: CPU frame time was roughly 9–12 ms
while the GPU needed only 3.8–3.9 ms. The follow-up preserves exact alpha order
and every quality/range setting while replacing comparison sorting with a
stable radix, avoiding a second immediate-buffer pack for deferred transparent
instances, packing per-instance light sets, caching static picking descriptors,
and using checked direct group handles. Historical material groups now retire
after one complete unused frame (not after one `Draw`, because landscape slices
and paperdoll share the dispatcher) and release their list capacity. That dense
sample reached roughly 5.3–6.2 ms CPU / 7.1–8.3 ms p95 and 3.2–3.6 ms GPU before
the final group-retirement review correction.
The earlier software limiter also used coarse `Thread.Sleep`, whose nominal
one-millisecond waits measured about 15.98 ms on this host. A reusable Windows
high-resolution waitable timer now paces to the monitor deadline and is reported
as `pace` by the frame profiler. Scripted UI automation no longer enables the
allocation-heavy GPU-stream dump implicitly, and graceful shutdown now wakes
all persistent mesh workers without resetting their shared wake signal. The
post-review connected stress rerun remains pending because local ACE was not
listening; sparse-scene 239 FPS samples are not substituted for that gate.
**Files:** `src/AcDream.App/Rendering/RetailAlphaQueue.cs`;
`src/AcDream.App/Rendering/Wb/WbDrawDispatcher.cs`;
`src/AcDream.App/Rendering/ParticleRenderer.cs`;
@ -129,7 +151,9 @@ used only for memory comparison. The lifestone/particle visual gate remains.
`src/AcDream.App/Rendering/CompositeTextureArrayCache.cs`;
`src/AcDream.App/Rendering/StandaloneBindlessTextureCache.cs`;
`src/AcDream.Content/BoundedDatObjectCache.cs`;
`src/AcDream.Content/DecodedTextureCache.cs`.
`src/AcDream.Content/DecodedTextureCache.cs`;
`src/AcDream.App/Rendering/FramePacingController.cs`;
`src/AcDream.App/Rendering/WindowsHighResolutionFramePacingWaiter.cs`.
**Research:**
`docs/research/2026-07-18-retail-shared-alpha-list-pseudocode.md`;