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

measure(perf): per-component ablation of the TunnelNetwork path; exact worms restored

Browse Source 
Round (Codex; reviewed by me against PERF-SAMPLED.md).

Revert: voxel.WormNoiseMode / voxel.WormLatticeStep, their storage and their switches are
REMOVED. VoxelWormField has one path again, exact Perlin with the N1 short circuit. Kept:
bEnableWorms (asset), WormsForceOff (measurement only), the exact block skip (default 0).

Ablation instrument: 14 development-only, world-changing force-off bits (voxel.TunnelAblate*),
applied on the fused, op-stack/classifier, cached SDF/core and modifier/floor paths. They resolve
once per process, are hashed into the generation fingerprint, and are compiled to zero in
UE_BUILD_SHIPPING (VoxelDensityAblation.cpp). They are a measurement instrument, never a gameplay
or multiplayer switch.

Defaults exact: canonical export OBJ b3e5f4c3..9b377, worms-region 311CA42C / 75,482 triangles,
capability 20,830 / 10,909, and 841/841 keyed tiles identical to NextHeadStaticTrace_20260913.

Static worker seconds saved by removing ONE component (two off runs vs the interleaved all-on
33.06 s; baseline noise 0.27%):
  tunnel SDF 11.4%, cave warp 10.3%, passage carving 10.1%, landing posts 9.9%,
  passage structural posts 5.6%, worms 2.4%, native floor 2.2%, origin spine 1.6%,
  room SDF 1.6%, everything else <= 1.1%
Moving (2+2 runs each) confirms the top five at 6.0-11.6%. Tunnel SDF off cuts LOD0 generation
p95 104 -> 83 ms: it owns the slow tail. The positive savings sum to 58% of the static total; the
other 42% is shared. The rows are not additive.

Unexplained: tunnel-core OFF made the static total 2.9x SLOWER (95.5 s) while LOD0 generation got
faster (p50 64.7 -> 60.6 ms) and triangles fell. The report blames the changed field, which does
not fit; more likely a slow fallback path in coarse tiles. That row is not a cost measurement.

Ledger: every Unreal process exited 0 and verified the loaded DLL; two PowerShell parser errors
happened before any process started. No crashes or retries.

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 09:39:57 +02:00
1 parent c132243b30
commit 6ccf3718ff
12 files changed
+833 -695
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PERF-SAMPLED.md
+237 -10
View File
@@ -1226,21 +1226,21 @@ Both runs also reported `canonical_repeat_equal=true`. The disabled asset path t
the hand-zeroed control byte-for-byte at the exported geometry hash, while exceeding the required
20,830 / 10,909 capability gate.
### Part 2 — measured cheap field
### Historical Part 2 — measured cheap field (reverted 2026-09-14)
The selected field is `voxel.WormNoiseMode=1`: a deterministic scalar value-noise evaluation
using the same worm coordinate mapping, threshold, N1 short-circuit, N2 offset, and strength
contract. The code default is mode 1. Mode 0 remains the legacy gradient-Perlin exact control.
`voxel.WormNoiseMode` and `voxel.WormLatticeStep` are documented development-only,
world-changing experiments; they are not the multiplayer off switch and must be identical on
all peers and on every regeneration of a world.
The historical build selected `voxel.WormNoiseMode=1`: a deterministic scalar value-noise
evaluation using the same worm coordinate mapping, threshold, N1 short-circuit, N2 offset, and
strength contract. That selection was reverted after the ablation review. The current source is
exact-only Perlin mode 0; `voxel.WormNoiseMode` and `voxel.WormLatticeStep`, their storage, and
their switches no longer exist. The run names and measurements below are retained as historical
evidence only.
The coarser lattice was measured rather than assumed. The first lattice implementation was much
slower because of cache churn; after the fixed tile-window cache, it returned to roughly the
exact cost but did not improve it. Representative static worker seconds were exact 33.10,
lattice-2 initial 160.86, lattice-2 cache-fixed 33.78, lattice-4 cache-fixed 33.25, and exact
block-skip 33.10. The lattice path is retained as a bounded experiment, but is not the selected
production mode. The exact block skip is also retained as a fail-closed exact-mode diagnostic;
block-skip 33.10. The lattice path was a bounded historical experiment and is not retained. The
exact block skip is retained as a fail-closed exact-mode diagnostic;
it is not used as a correctness or performance claim for value noise because its Perlin proof
does not cover the changed field.
@@ -1255,7 +1255,7 @@ The capability and geometry candidates were:
| `WormCheapValueLattice2Explore_20260914` | value, lattice 2 | `C47567A2` | 75,132 | 20,857 / 10,917 | pass; no speed gain |
| `WormCheapFastHashExplore_20260914` | value, lattice 0 | `6204C7E6` | 75,956 | 20,851 / 10,917 | selected |
The selected mode changes the field, but the measured full-game gain is small rather than a
The historical selected mode changed the field, but the measured full-game gain was small rather than a
claim of a dramatic speedup. In two interleaved static pairs, exact averaged 0.087128 / 0.148415
s request-to-ready p50/p95, 0.063042 / 0.105080 s generation p50/p95, and 33.213871 worker s;
cheap averaged 0.087761 / 0.145897 s, 0.062863 / 0.106258 s, and 33.083774 worker s. In two
@@ -1438,3 +1438,230 @@ capability or comparison gate.
| `WormFinalForceOffBuildSamplerOfflineRetry_20260914` | PASS; retried with writable `Saved\ZenData`/`Saved\DDC`, 733 modules, 10 loaded, zero identity mismatches, all 7,795 unique PCs mapped. |
| `PerfSampledBuild_FinalFingerprint.log` | PASS; final cache-key hardening rebuild, Result: Succeeded. |
| `WormFinalFingerprintSmoke_20260914` | PASS; post-hardening clean game smoke, exact staged runtime verified, writable DDC, zero crash folders/reporters. |
## Cheap-worm revert and TunnelNetwork ablation round — 2026-09-14
This section supersedes the historical cheap-worm selection above. The earlier value-noise
experiment changed the field without producing a gain: static movement was within noise and the
moving worker result was about 2.1% slower. The production/default path therefore keeps exact
Perlin.
### Revert and exactness
- voxel.WormNoiseMode and voxel.WormLatticeStep, their storage, and their switches are removed.
- VoxelWormField now has one evaluation path: exact Perlin, with the existing N1 short circuit.
- The exact block skip remains available as voxel.WormBlockSkip, default 0. The asset-owned
bEnableWorms and the measurement-only voxel.WormsForceOff remain.
- The generation fingerprint still hashes bEnableWorms and WormsForceOff, and now also hashes the
resolved TunnelNetwork ablation mask.
- The new ablation controls are development-only, resolve once per process, are world-changing,
and are never gameplay or multiplayer switches. A nonzero bit means force that stage off; zero
means all stages on. Shipping builds resolve the mask to zero.
The prescribed UBT command from this file was used after synchronizing the staged host copy under
the plugin Saved directory. The final build log is
E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\PerfSampledBuild_Ablation.log and reports
Result: Succeeded. The loaded runtime was
E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\BuildHost\HostProject\Plugins\VoxelForge\Binaries\Win64\UnrealEditor-VoxelForge.dll
with SHA-256 A70528AC82641D7D610D8BA9497D0322B1EBDE22978B1E826A97FC4C0D038149. The editor module
SHA-256 was 8EA2A71BB29F828C5DE7559EC200AC4E86CF61F7DB5C8470406208073E22456D.
The two default commandlets both used
E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\DDC and
E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\ZenData:
- AblationDefaultExact_20260914Out: mask 0; status ok; canonical 128^3 export SHA-256
B3E5F4C398DD0547C6C55F415872058246D4A292C2F0D636EE1CD0829DE9B377; geometry CRC 07C14005;
82,273 vertices and 153,346 triangles.
- AblationDefaultCapability_20260914Out: mask 0; geometry CRC 311CA42C; 75,482 triangles;
player_fit_volume_cells 20,830; reachable_player_fit_cells 10,909; walk-only reachable
10,909; Connected; deterministic JSON.
The current game startup trace
E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\AblationDefaultStaticTrace_20260914.startup.jsonl
was compared by tile key to
E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\NextHeadStaticTrace_20260913.startup.jsonl:
841 versus 841 tile records, zero missing, zero extra, zero verdict/empty/triangle mismatches,
and identical all-tile and LOD0 triangle totals of 669,834 and 325,204. The current trace
finished with desired_set_satisfied_and_queue_drained and the loaded runtime hash above.
### Ablation controls
| bit | CVar | stage forced off |
|---:|---|---|
| 0x0001 | voxel.TunnelAblateCaveWarp | the three cave-warp Perlin calls; CaveWarpStrength resolves to zero |
| 0x0002 | voxel.TunnelAblateDetailOps | detail noise, roughness, and terrain operations |
| 0x0004 | voxel.TunnelAblateRoomSDF | room SDF and room-SDF joins in EvaluateSDFCached |
| 0x0008 | voxel.TunnelAblateTunnelSDF | tunnel SDF in EvaluateSDFCached |
| 0x0010 | voxel.TunnelAblateTunnelCore | EvaluateTunnelCoreWorld, support/core SDF work, and its GetDensityAt tail |
| 0x0020 | voxel.TunnelAblatePassageCarving | EvaluateModifierSDF and passage carving |
| 0x0040 | voxel.TunnelAblatePassageStructuralPosts | passage structural posts and legacy structural floor support |
| 0x0080 | voxel.TunnelAblateNativeFloor | native floor composition |
| 0x0100 | voxel.TunnelAblateDisturbances | ApplyDisturbances |
| 0x0200 | voxel.TunnelAblateOriginSpine | origin spine |
| 0x0400 | voxel.TunnelAblateBoundarySeal | boundary seal |
| 0x0800 | voxel.TunnelAblateLandingPosts | landing posts and landing-floor posts |
| 0x1000 | voxel.TunnelAblateXYEdgeSeal | XY edge seal |
| 0x2000 | voxel.TunnelAblatePitChimneySDF | pit and chimney SDF work |
Each override is applied in the fused evaluator, operator-stack source/classifier paths, cached
SDF/core paths, and modifier/floor paths where that stage exists. The resolved mask is emitted in
the explore JSON and is part of the generation fingerprint. This is an ablation instrument, not a
candidate field.
### Measurement method and noise
All game runs were sampler-off, NullRHI, operator block on, fused evaluator on, outer classifier
off, and used a 15-second clean session. Static runs produced 343 LOD0 samples and 841 applied
tiles. Each static row has two off runs; the two all-on static baselines are shared and interleaved
across the two passes. The all-on static mean was 33.064458 worker seconds, generation p50
0.064657 s, generation p95 0.105786 s. Baseline worker range was 0.27%; baseline generation-p95
range was 1.62%.
The table reports the mean of the two off runs, the saving against that all-on mean, and the
off-run worker range as a compact noise indicator. Positive saving means the off run used fewer
worker seconds. All rows completed 841 tiles with zero obsolete aborts.
| component removed | runs | off worker s | saving s (%) | off generation p50 s | off generation p95 s | off worker range |
|---|---:|---:|---:|---:|---:|---:|
| tunnel SDF | 2 | 29.304 | 3.760 (11.37%) | 0.060813 | 0.084811 | 0.53% |
| cave warp | 2 | 29.654 | 3.411 (10.31%) | 0.056881 | 0.098260 | 2.03% |
| passage carving | 2 | 29.730 | 3.334 (10.08%) | 0.056336 | 0.096892 | 0.88% |
| landing posts | 2 | 29.800 | 3.264 (9.87%) | 0.056776 | 0.097884 | 1.33% |
| passage structural posts | 2 | 31.205 | 1.859 (5.62%) | 0.060143 | 0.099723 | 0.50% |
| worms | 2 | 32.265 | 0.799 (2.42%) | 0.064232 | 0.099511 | 0.83% |
| native floor | 2 | 32.347 | 0.718 (2.17%) | 0.063590 | 0.103897 | 0.45% |
| origin spine | 2 | 32.529 | 0.535 (1.62%) | 0.063174 | 0.102446 | 0.61% |
| room SDF | 2 | 32.542 | 0.522 (1.58%) | 0.064194 | 0.100999 | 0.90% |
| pit/chimney SDF | 2 | 32.689 | 0.375 (1.13%) | 0.064599 | 0.103472 | 0.26% |
| boundary seal | 2 | 32.814 | 0.251 (0.76%) | 0.064863 | 0.104105 | 1.21% |
| detail ops | 2 | 32.838 | 0.227 (0.69%) | 0.065330 | 0.100864 | 0.69% |
| XY edge seal | 2 | 32.901 | 0.164 (0.49%) | 0.065021 | 0.104411 | 1.30% |
| disturbances | 2 | 33.208 | -0.144 (-0.43%) | 0.065100 | 0.106342 | 0.53% |
| tunnel core | 2 | 95.468 | -62.404 (-188.73%) | 0.060597 | 0.098965 | 0.07% |
The field was intentionally allowed to change. For example, tunnel-SDF off produced 643,388
triangles, cave-warp off 655,726, passage-carving off 654,932, and tunnel-core off 655,654,
against the all-on 669,834 static triangles. Some low-cost rows retained the same mesh totals in
this window; they are still world-changing overrides and are not correctness candidates.
The sum of positive standalone static savings is 19.218608 worker seconds, 58.125% of the
33.064458-second baseline, leaving 13.845850 seconds (41.875%) as shared/unattributed cost in
this ablation accounting. The signed sum over every row is -43.329105 seconds because tunnel-core
removal is a large regression caused by the changed field; the rows are not additive and this is
why the signed sum is not a speedup prediction.
Moving confirmation covered the five largest positive static savers. Each row has two fresh
all-on baselines interleaved with two off runs. Movement began at 3 seconds and ran at 800 cm/s;
each run produced 882 LOD0 samples and 1,680 applied tiles, with zero obsolete aborts.
| component removed | runs | all-on worker s | off worker s | saving s (%) | all-on gen p50/p95 s | off gen p50/p95 s | baseline/off worker range |
|---|---:|---:|---:|---:|---:|---:|---:|
| tunnel SDF | 2+2 | 72.218 | 63.847 | 8.371 (11.59%) | 0.063238 / 0.104331 | 0.060117 / 0.082540 | 0.71% / 0.87% |
| cave warp | 2+2 | 72.810 | 65.010 | 7.800 (10.71%) | 0.064610 / 0.105986 | 0.055398 / 0.094423 | 0.46% / 1.14% |
| landing posts | 2+2 | 72.556 | 64.763 | 7.794 (10.74%) | 0.063820 / 0.105220 | 0.055539 / 0.096228 | 1.13% / 2.36% |
| passage carving | 2+2 | 72.043 | 64.552 | 7.492 (10.40%) | 0.063507 / 0.104761 | 0.055276 / 0.096373 | 1.07% / 1.72% |
| passage structural posts | 2+2 | 72.417 | 68.093 | 4.324 (5.97%) | 0.063779 / 0.104611 | 0.059068 / 0.101321 | 0.14% / 0.83% |
The moving standalone saving sum is 35.780708 worker seconds. Because these are five separate
ablations with separate baseline pairs, it is not a combined-speedup forecast; against the mean
72.409-second baseline per row it is 49.4%, leaving 36.628 seconds of shared/unattributed work.
### Ranking and plausible cheaper forms
1. Tunnel SDF — the largest measured real cost in both paths. A plausible exact optimization is
sharing one cached swept-tunnel result between EvaluateSDFCached, tunnel-core metadata, and
the final density tail, while preserving the existing evaluation order and float values. That
would be field-preserving. A lower-resolution or analytic approximation would be field-changing
and would need a fresh capability/export gate.
2. Cave warp — the three Perlin calls are expensive in the ablation. Exact reuse of the same
coordinates/results across the evaluator, cached graph, and op-stack paths is plausible and
field-preserving. Reducing frequency or replacing the noise is field-changing.
3. Passage carving — exact memoization/reuse of modifier SDF and passage candidate results across
EvaluateModifierSDF, ApplyPassageCarvingOnly, and the structural/floor queries is plausible and
field-preserving. Simplifying the carving geometry is field-changing.
4. Landing posts — exact per-tile reuse of landing/post/floor metadata is plausible and
field-preserving. Removing or thinning landing posts is field-changing.
5. Passage structural posts — exact reuse of structural-post SDF/metadata is plausible and
field-preserving. Removing or merging posts is field-changing.
Worms is sixth at 2.42% static and is now kept exact because the cheap alternatives had no gain.
Native floor and origin spine follow at 2.17% and 1.62%. Room SDF, detail ops, seals,
disturbances, and pit/chimney work are not attractive next-round targets on these measurements.
Tunnel-core removal is explicitly not a candidate: it increased worker time by about 2.9x.
These ablations explain why the old 19.4% line-level worm figure and the proposed tunnel-core
line targets were unsafe attribution guides.
### Complete launch ledger for this round
All game and commandlet logs below are under
E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved. Every commandlet used the writable DDC path
E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\DDC. No offline symbolizer was needed or
launched in this round. Every Unreal process that ran exited 0, verified the loaded runtime
SHA-256 A70528AC82641D7D610D8BA9497D0322B1EBDE22978B1E826A97FC4C0D038149, and had no crash or
retry; the two parser-only attempts below launched no Unreal process. Peak observed private
memory was approximately 1.54–1.60 GiB per process, with no runaway growth.
| launch | outcome |
|---|---|
| initial prescribed staged build | UP-TO-DATE report; staged host was stale, so no measurement was taken |
| PerfSampledBuild_Ablation | PASS; prescribed UBT build, Result: Succeeded |
| AblationDefaultExact_20260914 | PASS; canonical export exact |
| AblationDefaultCapability_20260914 | PASS; 311CA42C, 75,482 triangles, 20,830 / 10,909 |
| static ablation batch launch 1 | FAILED before process start; PowerShell parser error from a missing DDC-argument quote |
| static ablation batch launch 2 | FAILED before process start; same quoting error; no Unreal process |
| AblationStaticBaseA_20260914 | PASS |
| AblationStaticBaseB_20260914 | PASS |
| AblationStaticCaveWarpA_20260914 | PASS |
| AblationStaticCaveWarpB_20260914 | PASS |
| AblationStaticDetailOpsA_20260914 | PASS |
| AblationStaticDetailOpsB_20260914 | PASS |
| AblationStaticRoomSDFA_20260914 | PASS |
| AblationStaticRoomSDFB_20260914 | PASS |
| AblationStaticTunnelSDFA_20260914 | PASS |
| AblationStaticTunnelSDFB_20260914 | PASS |
| AblationStaticTunnelCoreA_20260914 | PASS; 95.436091 worker seconds |
| AblationStaticTunnelCoreB_20260914 | PASS; 95.500636 worker seconds |
| AblationStaticPassageCarvingA_20260914 | PASS |
| AblationStaticPassageCarvingB_20260914 | PASS |
| AblationStaticPassageStructuralPostsA_20260914 | PASS |
| AblationStaticPassageStructuralPostsB_20260914 | PASS |
| AblationStaticNativeFloorA_20260914 | PASS |
| AblationStaticNativeFloorB_20260914 | PASS |
| AblationStaticDisturbancesA_20260914 | PASS |
| AblationStaticDisturbancesB_20260914 | PASS |
| AblationStaticOriginSpineA_20260914 | PASS |
| AblationStaticOriginSpineB_20260914 | PASS |
| AblationStaticBoundarySealA_20260914 | PASS |
| AblationStaticBoundarySealB_20260914 | PASS |
| AblationStaticLandingPostsA_20260914 | PASS |
| AblationStaticLandingPostsB_20260914 | PASS |
| AblationStaticXYEdgeSealA_20260914 | PASS |
| AblationStaticXYEdgeSealB_20260914 | PASS |
| AblationStaticPitChimneySDFA_20260914 | PASS |
| AblationStaticPitChimneySDFB_20260914 | PASS |
| AblationStaticWormsOffA_20260914 | PASS |
| AblationStaticWormsOffB_20260914 | PASS |
| AblationMovingTunnelSDFBaseA_20260914 | PASS |
| AblationMovingTunnelSDFA_20260914 | PASS |
| AblationMovingTunnelSDFBaseB_20260914 | PASS |
| AblationMovingTunnelSDFB_20260914 | PASS |
| AblationMovingCaveWarpBaseA_20260914 | PASS |
| AblationMovingCaveWarpA_20260914 | PASS |
| AblationMovingCaveWarpBaseB_20260914 | PASS |
| AblationMovingCaveWarpB_20260914 | PASS |
| AblationMovingPassageCarvingBaseA_20260914 | PASS |
| AblationMovingPassageCarvingA_20260914 | PASS |
| AblationMovingPassageCarvingBaseB_20260914 | PASS |
| AblationMovingPassageCarvingB_20260914 | PASS |
| AblationMovingLandingPostsBaseA_20260914 | PASS |
| AblationMovingLandingPostsA_20260914 | PASS |
| AblationMovingLandingPostsBaseB_20260914 | PASS |
| AblationMovingLandingPostsB_20260914 | PASS |
| AblationMovingPassageStructuralPostsBaseA_20260914 | PASS |
| AblationMovingPassageStructuralPostsA_20260914 | PASS |
| AblationMovingPassageStructuralPostsBaseB_20260914 | PASS |
| AblationMovingPassageStructuralPostsB_20260914 | PASS |
| AblationDefaultStaticTrace_20260914 | PASS; current 841-tile trace, exact keyed comparison |
No .uasset was edited. Engine source and binaries were not modified. No commit, push, or stash was
performed.