Fr0zka/VoxelForge
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

perf(generation): exact worm block skip with a proven 3D Perlin bound -- measured NEUTRAL

Browse Source 
The sampler put the worm carve's first noise call (VoxelGenerator.cpp:3985, FusedWorm) at 19.4% of
LOD0. The worm only carves where N1 + N2 < T with N2 >= 0, so where N1 >= T it contributes nothing,
bit for bit. A block proven to have N1 >= T everywhere can skip both noise calls exactly.

- The Lipschitz bound for this Perlin3D implementation: L = (15/4) * sqrt(3) * 1.25 = 8.118988160,
  in scaled-N1 units, with a 1e-3 float-rounding margin. Adversarial sampling found a worst ratio
  of 3.187 (0.39 L). The existing 2D floor-relief bound (8.5) is a component bound of a different
  field and was confirmed not to transfer.
- A soundness test that can fail (VoxelForgeWormBlockSkipTest): 6,476 proved blocks, 355,945 exact
  lattice samples, 0 violations. A deliberately halved L proves a bad block; the full L rejects it.
- voxel.WormBlockSkip, default 1.

Field exact: export OBJ byte-identical (b3e5f4c3..9b377), 841/841 keyed tiles identical to
NextHeadStaticTrace_20260913, capability 20,830 / 10,909.

RESULT, stated plainly: no measurable speed gain. The skip rate is 4.07% of worm-eligible LOD0
samples and 0% at LOD1-4, because the bound is ~2.5x looser than the worst observed ratio at the
block size used. Paired on-vs-off, two interleaved runs each: static worker +2.15%, generation p50
+1.8%; moving worker -1.42%, request p50 -2.5%. Within noise. The next round turns the default OFF.
The proven bound and its test are kept for a tighter skip later (smaller or hierarchical blocks,
or a per-cell bound).

Run ledger (from Codex's report): the first build failed on an ambiguous test overload and was
fixed; two timing runs are void because the orchestrator stopped the session so the owner could
use the PC; the eight valid timing runs, the sampler run and the diagnostics runs completed. No
new crash folders. PERF-SAMPLED.md now lists all eight earlier-round UECC crash records.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y
This commit is contained in:
Fr0zkaandClaude Opus 5 committed 2026-09-14 04:29:11 +02:00
1 parent 5489b5e311
commit 26b34765d5
9 files changed
+987 -24
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PERF-SAMPLED.md
+172
View File
@@ -928,3 +928,175 @@ Smaller unresolved groups map to `UnrealEditor-Engine.dll + RVA 0x500947` (gener
This closes the actionable part of candidate 5: the previous plugin constructor was removed, while
the remaining unknown is an engine-symbolisation limitation. No commit, push, or stash was performed.
## Worm block skip round (2026-09-14)
### Proof and gates
The worm skip is enabled by default through `voxel.WormBlockSkip=1`. It is fail-closed and only
removes the two worm noise calls after the unchanged `WormNetworkRange` mask has passed. The
worker-local cache is keyed by generator owner, seed, params/layout fingerprint, tile origin,
step, lattice size, and worm parameters; it is never shared across tiles or workers.
For this implementation's actual `GradDot` set, the unique gradients are
`(±1,±1,0)`, `(±1,0,±1)`, and `(0,±1,±1)`. The exact finite-support partial-derivative relaxation
with `Fade'(t)=30t²(1-t)²` has supremum `15/4`, attained at `(0.5,0.5,0.5)`. Therefore the
noise-space vector bound is
`L = (15/4) * sqrt(3) * VOXEL_NOISE_SCALE = 8.118988160` (`VOXEL_NOISE_SCALE=1.25`).
The existing 2D floor constant was proven differently and does not transfer: its comment uses
the loose component bound `[-1,1]`, four-edge interpolation difference bound, and
`max SmoothCurve' = 1.875`, giving `1 + 4*1.875 = 8.5` per partial. The worm derives its
bound from the actual 3D gradient support. The block radius is measured after the exact X/Y/Z
noise transform, including `WormFrequency`, `VerticalScale`, and `WormHorizontalBias`; `Abs` is
1-Lipschitz and the positive noise scale is included in `L`. A `1e-3` scaled-output margin is
subtracted before the threshold comparison. It is deliberately much larger than the accumulated
float rounding at the center/radius operations and is covered by the soundness test.
The adversarial attack used dense corner/face sampling plus random pairs: 17,365,280 pairs,
worst scaled ratio `3.187330411`, or `0.392577295*L`. The fail-capable automation test also
audited 6,476 proved blocks over 355,945 exact lattice samples with zero violations. A deliberately
halved `L` proves the negative-control block while the exact point has `N1=0.011259466 < 0.015`
and the full `L` rejects it.
Default-on export is byte-identical to the baseline: geometry hash
`B3E5F4C398DD0547C6C55F415872058246D4A292C2F0D636EE1CD0829DE9B377`, 82,273 vertices,
153,346 triangles. The static trace is 841/841 keyed tiles with zero field/empty/triangle
differences. Capability remains 20,830 / 10,909.
### Clean game A/B
Sampler off, diagnostics off, `voxel.OuterClassifierMode=0`, interleaved static and moving runs.
Static runs had 343 LOD0 ready samples; moving runs had 1,274. All valid runs had identical
applied geometry totals, zero validation calls, zero obsolete worker tasks, and verified the loaded
plugin module. The clean timings below used the staged runtime before the final counter-only
diagnostic-accounting rebuild (`81F52389A78543E41EEC8E3E40DEF8F2634A79EB83E2A933FF35EE42CD611E12`).
That patch is behind diagnostics, so it does not change this clean path. The final staged runtime
used by the last diagnostic run is `E72E140100B3D258F28E9DBA19CD95D379D98794DA49D24C708FDD01B5B5FE08`.
| run | switch | request p50/p95 (s) | generation p50/p95 (s) | worker s |
|---|---|---:|---:|---:|
| WormSkipGameOffStaticA_20260914 | off | 0.087151 / 0.150116 | 0.063115 / 0.104426 | 33.308914 |
| WormSkipGameOnStaticA_Resume_20260914 | on | 0.092174 / 0.154458 | 0.065730 / 0.110507 | 34.980771 |
| WormSkipGameOffMovingA_20260914 | off | 0.066421 / 0.138179 | 0.059918 / 0.108181 | 99.193222 |
| WormSkipGameOnMovingA_20260914 | on | 0.064163 / 0.135519 | 0.059273 / 0.106991 | 97.624292 |
| WormSkipGameOffStaticB_20260914 | off | 0.090372 / 0.160169 | 0.064016 / 0.109362 | 34.116813 |
| WormSkipGameOnStaticB_Resume_20260914 | on | 0.091922 / 0.154718 | 0.063731 / 0.112871 | 33.892686 |
| WormSkipGameOffMovingB_Resume_20260914 | off | 0.066659 / 0.142168 | 0.061916 / 0.110043 | 100.995082 |
| WormSkipGameOnMovingB_Resume_20260914 | on | 0.065557 / 0.141592 | 0.060515 / 0.110629 | 99.716139 |
Paired means, on relative to off:
| mode | request p50/p95 | generation p50/p95 | worker s | on vs off worker |
|---|---:|---:|---:|---:|
| static | off 0.088762 / 0.155143; on 0.092048 / 0.154588 | off 0.063566 / 0.106894; on 0.064731 / 0.111689 | 33.712864 → 34.436729 | +2.1471% |
| moving | off 0.066540 / 0.140174; on 0.064860 / 0.138556 | off 0.060917 / 0.109112; on 0.059894 / 0.108810 | 100.094152 → 98.670216 | −1.4226% |
The two-run result is mixed: moving improves modestly, while static is within ordinary run noise
and is slightly worse on worker sum. The switch remains default-on because the field proof is exact;
the timing result is not presented as a universal speed claim.
### Skip rate
The final diagnostics-enabled moving run used the final DLL and emitted 2,360 tile profiles. The
rate is `worm_block_skipped / worm_eligible`, not a fraction of all density samples:
| LOD | worm-eligible samples | successful block proofs | skipped eligible samples | skip rate |
|---:|---:|---:|---:|---:|
| 0 | 12,381,885 | 9,541 | 503,696 | 4.068007% |
| 1 | 4,821,000 | 0 | 0 | 0.000000% |
| 2 | 188,063 | 0 | 0 | 0.000000% |
| 3 | 58,276 | 0 | 0 | 0.000000% |
| 4 | 25,253 | 0 | 0 | 0.000000% |
### Sampler-on moving profile
`WormSkipGameOnMovingSampler_20260914` retained 96,649 samples with zero drops and zero capture
failures: 61,869 LOD0 and 34,780 LOD1+. It ran for 24.972305 s; resolved-frame fraction was
52.9174%, unknown-frame fraction 47.0826%, wait-like fraction 0.1011%, and resolved-leaf
fraction 95.3388%.
LOD0 exclusive top 20:
| # | samples | % | leaf |
|---:|---:|---:|---|
| 1 | 12472 | 20.1587 | `GetDensityWithParams` — `VoxelGenerator.cpp:3128` |
| 2 | 3137 | 5.0704 | `GetDensityAt` — `VoxelGenerator.cpp:2994` |
| 3 | 2853 | 4.6114 | unknown |
| 4 | 2540 | 4.1054 | `Sqrt` |
| 5 | 2238 | 3.6173 | `VoxelSDF::SmoothMin` — `VoxelCaveMorphology.h:170` |
| 6 | 1736 | 2.8059 | `VF_ProjectNativePassageFloorUncached` — `VoxelStrateManager.cpp:561` |
| 7 | 1704 | 2.7542 | `FChunkSDFSpatialIndex::GetRange` — `VoxelCaveMorphology.cpp:1356` |
| 8 | 1434 | 2.3178 | `VF_ProjectTunnelSegmentXY` — `VoxelCaveMorphology.cpp:970` |
| 9 | 1244 | 2.0107 | `ProjectWalkableTunnelFloor` — `VoxelPassageGeometry.h:338` |
| 10 | 1229 | 1.9865 | `VF_EvaluateSweptTunnelChain` — `VoxelCaveMorphology.cpp:1128` |
| 11 | 1151 | 1.8604 | `EvaluateTunnelCoreWorld` lambda — `VoxelCaveMorphology.cpp:7635` |
| 12 | 1009 | 1.6309 | swept-chain lambda — `VoxelCaveMorphology.cpp:1074` |
| 13 | 938 | 1.5161 | `EvaluateSDFCached` lambda — `VoxelCaveMorphology.cpp:7526` |
| 14 | 819 | 1.3238 | `IsFiniteFast` — `VoxelTypes.h:30` |
| 15 | 769 | 1.2429 | `IsFinite` — `VoxelTypes.h:47` |
| 16 | 764 | 1.2349 | `EvaluateModifierSDF` — `VoxelStrateManager.cpp:2451` |
| 17 | 748 | 1.2090 | `VF_EvaluatePassageLandingSDF` — `VoxelCaveMorphology.cpp:5673` |
| 18 | 739 | 1.1945 | `GenerateMesh` — `VoxelMarchingCubesMesher.cpp:578` |
| 19 | 724 | 1.1702 | `FScopedTimer::End` — `VoxelDensityProfile.h:435` |
| 20 | 700 | 1.1314 | `VF_DistanceSquaredToAabb` — `VoxelCaveMorphology.cpp:74` |
LOD0 inclusive top 20:
| # | samples | % | inclusive function |
|---:|---:|---:|---|
| 1 | 61869 | 100.0000 | `Invoke` |
| 2 | 61869 | 100.0000 | `LowLevelTasks::FTask::Init` lambda |
| 3 | 61869 | 100.0000 | `LowLevelTasks::TTaskDelegate::Call` |
| 4 | 61869 | 100.0000 | `UE::Tasks::Private::FTaskBase::Init` lambda |
| 5 | 61869 | 100.0000 | `LowLevelTasks::TTaskDelegate::CallAndMove` |
| 6 | 61869 | 100.0000 | `UE::Tasks::Private::FTaskBase::TryExecuteTask` |
| 7 | 61869 | 100.0000 | `TExecutableTaskBase::ExecuteTask` |
| 8 | 61869 | 100.0000 | `AVoxelWorld::LoadTile` lambda — `VoxelWorld.cpp:3782` |
| 9 | 61869 | 100.0000 | `AVoxelWorld::GenerateTileResult` — `VoxelWorld.cpp:4142` |
| 10 | 61774 | 99.8464 | `GenerateMesh` — `VoxelMarchingCubesMesher.cpp:578` |
| 11 | 60310 | 97.4802 | `GetDensityAt` — `VoxelGenerator.cpp:2606` |
| 12 | 38073 | 61.5381 | `GetDensityWithParams` — `VoxelGenerator.cpp:3426` |
| 13 | 15666 | 25.3212 | `VF_ForEachSpatialCandidate` — `VoxelCaveMorphology.cpp:1570` |
| 14 | 9775 | 15.7995 | `EvaluateSDFCached` — `VoxelCaveMorphology.cpp:7484` |
| 15 | 7846 | 12.6816 | `VF_EvaluateSweptTunnel` — `VoxelCaveMorphology.cpp:1167` |
| 16 | 7460 | 12.0577 | `VF_EvaluateSweptTunnelChain` — `VoxelCaveMorphology.cpp:1128` |
| 17 | 7276 | 11.7603 | `ApplyPassageCarvingOnly` — `VoxelStrateManager.cpp:2596` |
| 18 | 7209 | 11.6520 | `EvaluateTunnelCoreWorld` — `VoxelCaveMorphology.cpp:7732` |
| 19 | 7128 | 11.5211 | `VF_ApplyPassageCarving` — `VoxelDensityPrimitives.h:75` |
| 20 | 7106 | 11.4856 | `EvaluateModifierSDF` — `VoxelStrateManager.cpp:2538` |
### Run ledger
Every worm-round launch is listed here, including invalidated launches and the compile retry:
| launch | outcome |
|---|---|
| Initial staged build | Failed compiling `VoxelForgeWormBlockSkipTest.cpp` because of an ambiguous `FMath::IsNearlyEqual` overload; fixed by an explicit cast. No Unreal process was launched. |
| Rebuilt staged plugin | Succeeded; final build log reports `Result: Succeeded`. |
| `VoxelForge.Determinism.WormBlockSkipSoundness` | Passed: 6,476 proofs / 355,945 lattice samples / 0 violations; halved-L control fails as intended. |
| `WormSkipFieldOn_20260914` | Passed; export byte-identical and loaded module verified. |
| `WormSkipStaticTrace_20260914` | Data passed 841/841; harness intentionally stopped the process after `.done` (`ExitCode=-1`), not a crash. |
| `WormSkipCapability_20260914` | Passed; capability 20,830 / 10,909. |
| `WormSkipGameOffStaticA_20260914` | Completed; valid off A. |
| `WormSkipGameOnStaticA_20260914` | Launched before the stop; void by owner instruction and excluded, even though a complete artifact exists. |
| `WormSkipGameOffMovingA_20260914` | Completed; valid on/off interleave partner. |
| `WormSkipGameOnMovingA_20260914` | Completed; valid on/off interleave partner. |
| `WormSkipGameOffStaticB_20260914` | Completed; valid off B. |
| `WormSkipGameOnStaticB_20260914` | Interrupted at the stop; partial module/RSS artifacts only, void and excluded. |
| `WormSkipGameOnStaticA_Resume_20260914` | Completed; valid on A replacement. |
| `WormSkipGameOffMovingB_Resume_20260914` | Completed; valid off B replacement. |
| `WormSkipGameOnMovingB_Resume_20260914` | Completed; valid on B replacement. |
| `WormSkipGameOnStaticB_Resume_20260914` | Completed; valid on B replacement. |
| `WormSkipGameOnMovingSampler_20260914` | Completed; sampler-on moving profile, zero drops/failures. |
| `WormSkipDiagnosticsMoving_20260914` | Completed; superseded by the final counter-semantics diagnostic run, not used for reported skip rates. |
| `WormSkipDiagnosticsMovingProofs_20260914` | Completed on final DLL; per-LOD skip rates above, module verified. |
The Saved crash inventory also contains earlier-round UECC artifacts, which are kept visible here:
three failed `Task4CurrentTracePIE` attempts on 2026-09-09 (plugin-manager/ShaderCore/DDC launch
failures), `GameMovingOn_20260913` (fatal shutdown), `MemoGameStaticOffA_20260913` (no writable
DDC graph; retried successfully), `TakeoverFinalMovingSamplerOnHost_20260913` (crash artifact;
`...Host2` retry succeeded), `Final3MovingSamplerOn_20260913` (fatal shutdown), and
`IsFiniteExact_20260913` (no writable DDC graph). No CrashReportClient crash occurred in the
worm-round launches above.