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

perf(generation): field-exact LOD0 overhead removal; the unknown bucket is Core.dll

Browse Source 
Field-preserving floor round (Codex; reviewed by me against the code):
- exact per-(sample, passage) caching of native/generic floor projections and walkable-air
  checks, behind voxel.FloorRound1PassageProjectionCache
- a zero-work disabled path for VoxelDensityProfile::FScopedTimer, unchanged when enabled
- pointer / range-for iteration in the proven-hot tunnel, passage, spatial and candidate loops
  (TArray RangeCheck was ~6.5% of LOD0 samples on the Development binaries the owner plays on)
- FVector2D temporaries replaced with scalar doubles, same operations in the same order
- FSurfaceColumn made trivial, so thread-local construction does no work. Verified safe:
  ComputeSurfaceColumn writes all five outputs unconditionally before any branch, and every read
  of Box.Cols is gated by Box.Computed

Verified: export OBJ byte-identical (b3e5f4c3..9b377); capability 20,830 / 10,909; all 8
interleaved game runs gave identical totals (841 tiles / 669,834 triangles static, 2,348 /
1,985,216 moving, the same as previous rounds). Per-tile identity for the always-on changes still
needs a check against a stored trace; that is queued first in the next round.

Game path (sampler off, two interleaved runs each):
  projection-cache switch alone: neutral (static worker +0.41%, moving -0.05%)
  vs the previous round (cfad3d1, same harness; cross-round numbers reproduced within ~1% here):
    LOD0 generation p50  static 76.6 -> 65.8 ms (-14%), moving 72.4 -> 59.5 ms (-18%)
    LOD0 request p50/p95 moving 79.4/158.5 -> 67.6/137.6 ms
    worker seconds       static 39.2 -> 33.4, moving 113.7 -> 97.9
  The always-on changes carry the gain; they have no switch, so this is a cross-round comparison.

The sampler's ~12.9% "unknown" LOD0 leaf is engine code: UnrealEditor-Core.dll + RVA 0xC080,
without a Core PDB. The old FSurfaceColumn label is gone. Next round names it by its callers.

PERF-SAMPLED.md: corrected a sentence that repeated my false "daytime use" explanation.

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 2026-09-13 21:00:34 +02:00
1 parent b839edd3a3
commit ab4bbe3680
9 files changed
+629 -170
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PERF-SAMPLED.md
+162
View File
@@ -766,3 +766,165 @@ No rank is assigned to the 74-field parameter copy: the sampler did not measure
## Handoff
This round changed only diagnostic instrumentation and report artifacts under E:\Projet Unreal\VoxelM\Plugins\VoxelForge. Engine source was not modified. No commit, push, or stash was performed. The key caveat is explicit: the instrument has good closure and a successful planted-cost proof, but its 1 ms game overhead is above the requested timing/worker-CPU noise on static and above timing noise on moving. Treat the large, repeated component differences as useful; do not treat small single-digit improvements as established until that overhead is brought inside the game baseline.
## Floor round 1 — field-preserving LOD0 work (2026-09-13)
This section is the final report for the subsequent LOD0 floor round. It is deliberately separate
from the field-changing world-space tunnel-core merge and cave warp work. No density expression was
changed, no terrain-op order was changed, and no engine source was modified.
### Changes
- Passage floor projection is memoised in `FPassageEvaluationCache::FloorProjections`, one indexed
entry per passage and worker. Native projection, generic fallback projection, and walkable-air
classification reuse the same float outputs for an exact `(sample, passage)` hit. The sample key
compares the original double XYZ values; cache invalidation includes manager lifetime, layout
version, and chunk. `voxel.FloorRound1PassageProjectionCache` disables only this cache for A/B.
- Disabled `VoxelDensityProfile::FScopedTimer` construction now performs the enabled/probe test and
carries only an inactive flag; bucket/token initialization and out-of-line destructor work are
absent. Enabled profiling still enters/exits the same scope and preserves `End()` early closure.
- Proven-hot room/tunnel/passage candidate loops cache counts and use `GetData()` pointers where the
preceding count or spatial-index range proves the access. Other array checks remain intact.
- The two exact floor-projection XY loops scalarise the existing double `TVector2` temporaries while
retaining the original float conversions and result casts. `FVector::DistSquared` float paths were
left alone.
- `FSurfaceColumn` is now a trivial payload. `ComputeSurfaceColumn` writes all five values before
publishing `Computed`, so removing never-read default values removes its generated TLS constructor
leaf without changing a column or density.
### Build and field/capability gates
The prescribed UBT build succeeded on the staged host. The loaded runtime was verified on every field,
capability, and game process as:
`E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\BuildHost\HostProject\Plugins\VoxelForge\Binaries\Win64\UnrealEditor-VoxelForge.dll`
SHA-256: `E54290C5056D63A12E407E39EED96935CCDC87F82882E9A607C84759DF8656D7`.
Field runs:
- `E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\BuildHost\HostProject\Plugins\VoxelForge\Saved\FloorR1BookField2Off_20260913Out`
- `E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\BuildHost\HostProject\Plugins\VoxelForge\Saved\FloorR1BookField2On_20260913Out`
Both reports are `status=ok`, deterministic repeat `passed`, geometry CRC `07C14005`, 82,273
vertices, and 153,346 triangles. Both `geometry.obj` files are byte-identical with SHA-256
`b3e5f4c398dd0547c6c55f415872058246d4a292c2f0d636ee1cd0829de9b377`. The standing 841/841
per-tile triangle-identity record is preserved; the new game runs also report identical static and
moving applied tile/visible-tile/triangle totals in every off/on case.
Capability run:
`E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\BuildHost\HostProject\Plugins\VoxelForge\Saved\FloorR1BookCapability_20260913Out\explore.json`
The report is `status=ok`, deterministic repeat `passed`, with `player_fit_volume_cells=20830`,
`reachable_player_fit_cells=10909`, and `walk_only_reachable_player_fit_cells=10909`.
### Clean game A/B
Sampler off, diagnostics off, LOD0, `voxel.OuterClassifierMode=0`, interleaved in the order
off/on static A, on/off moving A, off/on static B, on/off moving B. Static runs had 343 LOD0 ready
samples; moving runs had 1,274.
| run | cache | request p50/p95 (s) | generation p50/p95 (s) | worker s | applied tiles / visible / triangles |
|---|---:|---:|---:|---:|---:|
| OffStaticA | 0 | 0.091413 / 0.145960 | 0.066545 / 0.108736 | 33.354203 | 841 / 447 / 669834 |
| OnStaticA | 1 | 0.091688 / 0.151787 | 0.066298 / 0.107866 | 33.549964 | 841 / 447 / 669834 |
| OnMovingA | 1 | 0.067281 / 0.138041 | 0.059686 / 0.108387 | 97.919700 | 2348 / 1055 / 1985216 |
| OffMovingA | 0 | 0.067659 / 0.135287 | 0.059316 / 0.107414 | 98.007684 | 2348 / 1055 / 1985216 |
| OffStaticB | 0 | 0.088911 / 0.146950 | 0.064988 / 0.104965 | 33.246204 | 841 / 447 / 669834 |
| OnStaticB | 1 | 0.090368 / 0.151585 | 0.065554 / 0.107661 | 33.323367 | 841 / 447 / 669834 |
| OnMovingB | 1 | 0.067540 / 0.139517 | 0.059545 / 0.108946 | 97.915109 | 2348 / 1055 / 1985216 |
| OffMovingB | 0 | 0.067745 / 0.137666 | 0.059707 / 0.108898 | 97.926095 | 2348 / 1055 / 1985216 |
Paired means, on relative to off:
| mode | off request p50/p95 | on request p50/p95 | off generation p50/p95 | on generation p50/p95 | off worker s | on worker s | on vs off worker |
|---|---:|---:|---:|---:|---:|---:|---:|
| static | 0.090162 / 0.146455 | 0.091028 / 0.151686 | 0.065767 / 0.106851 | 0.065926 / 0.107764 | 33.300204 | 33.436666 | +0.41% |
| moving | 0.067702 / 0.136477 | 0.067411 / 0.138779 | 0.059512 / 0.108156 | 0.059616 / 0.108667 | 97.966890 | 97.917405 | −0.05% |
The isolated projection-cache A/B is therefore neutral within run-to-run noise (an earlier brief
blamed cross-round drift on "daytime use"; that was false, the machine was idle): static is
slightly worse, moving worker sum is effectively unchanged. The implementation is retained because
the reuse is exact and the switch attributes the hypothesis; the result is not claimed as a speed
win. All eight runs had `validation_density_calls=0`, `obsolete_worker_tasks=0`, and loaded-DLL
verification true. The one-second process monitor observed game RSS from 1,946,759,168 to
2,019,475,456 bytes; these are peak samples, not a cross-round memory baseline.
### Final sampler-on moving run
Raw samples:
`E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\BuildHost\HostProject\Plugins\VoxelForge\Saved\VoxelStackSamples_game_3352.csv`
Summary:
`E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\BuildHost\HostProject\Plugins\VoxelForge\Saved\VoxelStackSummary_game_3352.txt`
The run retained 102,854 samples, dropped 0, had 67,187 LOD0 samples, ran 24.978248 s, and had
unknown-frame fraction 0.472158. It reported 2,348 applied tiles, 1,055 visible tiles, 1,985,216
triangles, 2,360 generation tasks, 103.016440 worker seconds, and zero obsolete worker tasks.
LOD0 exclusive top 20:
| # | samples | % | leaf |
|---:|---:|---:|---|
| 1 | 11531 | 17.1625 | `GetDensityWithParams` — `VoxelGenerator.cpp:2874` |
| 2 | 8658 | 12.8864 | unknown |
| 3 | 3068 | 4.5664 | `GetDensityAt` — `VoxelGenerator.cpp:1998` |
| 4 | 2702 | 4.0216 | `Sqrt` |
| 5 | 2312 | 3.4411 | `VoxelSDF::SmoothMin` |
| 6 | 2209 | 3.2878 | `EvaluateTunnelCoreWorld` lambda — `VoxelCaveMorphology.cpp:7648` |
| 7 | 1876 | 2.7922 | `EvaluateSDFCached` lambda — `VoxelCaveMorphology.cpp:7526` |
| 8 | 1788 | 2.6612 | `VF_ProjectNativePassageFloorUncached` — `VoxelStrateManager.cpp:597` |
| 9 | 1504 | 2.2385 | `VF_DistanceSquaredToAabb` — `VoxelCaveMorphology.cpp:68` |
| 10 | 1419 | 2.1120 | `VF_ProjectTunnelSegmentXY` — `VoxelCaveMorphology.cpp:970` |
| 11 | 1268 | 1.8873 | `ProjectWalkableTunnelFloor` — `VoxelPassageGeometry.h:314` |
| 12 | 1241 | 1.8471 | `VF_EvaluateSweptTunnelChain` — `VoxelCaveMorphology.cpp:1006` |
| 13 | 1009 | 1.5018 | swept-chain lambda — `VoxelCaveMorphology.cpp:1076` |
| 14 | 880 | 1.3098 | `VF_EvaluatePassageLandingSDF` — `VoxelCaveMorphology.cpp:5690` |
| 15 | 844 | 1.2562 | `GenerateMesh` — `VoxelMarchingCubesMesher.cpp:578` |
| 16 | 758 | 1.1282 | `FScopedTimer::End` — `VoxelDensityProfile.h:432` |
| 17 | 732 | 1.0895 | `EvaluateModifierSDF` — `VoxelStrateManager.cpp:2470` |
| 18 | 729 | 1.0850 | `ApplyPassageStructuralPostsMC` — `VoxelStrateManager.cpp:2808` |
| 19 | 700 | 1.0419 | `VF_GetPassageFloorProjectionCacheEntry` — `VoxelStrateManager.cpp:349` |
| 20 | 609 | 0.9064 | `EvaluateSDFCached` lambda — `VoxelCaveMorphology.cpp:7397` |
LOD0 inclusive top 20:
| # | samples | % | inclusive function |
|---:|---:|---:|---|
| 1 | 67187 | 100.0000 | `Invoke` |
| 2 | 67187 | 100.0000 | `LowLevelTasks::FTask::Init` lambda |
| 3 | 67187 | 100.0000 | `LowLevelTasks::TTaskDelegate::Call` |
| 4 | 67187 | 100.0000 | `UE::Tasks::Private::FTaskBase::Init` lambda |
| 5 | 67187 | 100.0000 | `LowLevelTasks::TTaskDelegate::CallAndMove` |
| 6 | 67187 | 100.0000 | `UE::Tasks::Private::FTaskBase::TryExecuteTask` |
| 7 | 67187 | 100.0000 | `TExecutableTaskBase::ExecuteTask` |
| 8 | 67187 | 100.0000 | `AVoxelWorld::LoadTile` lambda — `VoxelWorld.cpp:3782` |
| 9 | 67186 | 99.9985 | `AVoxelWorld::GenerateTileResult` — `VoxelWorld.cpp:4126` |
| 10 | 67100 | 99.8705 | `UVoxelMarchingCubesMesher::GenerateMesh` — `VoxelMarchingCubesMesher.cpp:578` |
| 11 | 65522 | 97.5218 | `UVoxelGenerator::GetDensityAt` — `VoxelGenerator.cpp:2297` |
| 12 | 40538 | 60.3361 | `UVoxelGenerator::GetDensityWithParams` — `VoxelGenerator.cpp:3117` |
| 13 | 18270 | 27.1928 | `VF_ForEachSpatialCandidate` — `VoxelCaveMorphology.cpp:1582` |
| 14 | 11093 | 16.5106 | `VoxelCaveMorphology::EvaluateSDFCached` — `VoxelCaveMorphology.cpp:7377` |
| 15 | 8829 | 13.1409 | `EvaluateSDFCached` lambda — `VoxelCaveMorphology.cpp:7526` |
| 16 | 8777 | 13.0635 | `VF_EvaluateSweptTunnel` — `VoxelCaveMorphology.cpp:1167` |
| 17 | 8581 | 12.7718 | `VoxelCaveMorphology::EvaluateTunnelCoreWorld` — `VoxelCaveMorphology.cpp:7732` |
| 18 | 8366 | 12.4518 | `VF_EvaluateSweptTunnelChain` — `VoxelCaveMorphology.cpp:1029` |
| 19 | 7811 | 11.6258 | `ApplyPassageCarvingOnly` — `VoxelStrateManager.cpp:2596` |
| 20 | 7682 | 11.4338 | `VF_ApplyPassageCarving` — `VoxelDensityPrimitives.h:75` |
### Unknown-source resolution
The old `FSurfaceColumn` constructor PCs are absent from this final sampler. The LOD0 unknown leaf
is 8,658 samples (12.8864%), but the raw PC cluster that accounts for the largest part is not plugin
code: `0x00007FF88382C080`, `...C08A`, `...C095`, `...C0A5`, and `...C0A9` map to the loaded
`UnrealEditor-Core.dll` at runtime RVA `0xC080`, inside its private `.pdata` function range
`0xB4D0–0xC32B`. The packaged Core binary has no PDB, so the source function name cannot be resolved.
Smaller unresolved groups map to `UnrealEditor-Engine.dll + RVA 0x500947` (generated reflection
`StaticStruct` code) and `UnrealEditor-CoreUObject.dll + RVA 0x76168`; none maps to
`Plugins\VoxelForge`.
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.