Export tasks interleave with CPU tile jobs on the existing pool. Slow-dispatch keys
are logged (>5 ms, with -Timeline).
My runs outside the sandbox (Step21.exe, marker removed first, interleaved, GPU with
-Timeline):
CPU 1.356 / 1.321 / 1.277 s
GPU 1.453 (marker miss, 0 GPU tiles) / 1.249 / 1.216 / 1.217 s (match)
Export now starts at 686 ms in run c (was ~850), but the whole view is unchanged
(1.216 s): from first input (721 ms) to ready (1,216 ms) the GPU path spends ~500 ms
wall for ~255 ms of GPU timestamps. In runs b/d the output pool itself began late
(~860-875 ms). Slow tiles: (0,-6,-5,lod1) 28.5 ms, (-4,-15,-3,lod0) 11.3 ms,
(-3,-13,-2,lod0) 8.2 ms, all ordinary near-detail tiles.
Audits: hashes w1/w8, walk 5,642/5,642, near 960 / far 1,114 0 diffs vs CPU and
WARP, thin 353, survival 6.13 GiB, ten captures as step 18.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Coverage: the broad missing patch on far tiles was GPU skirt depth using the LOD key
step instead of the exported coarse mesher step (800 vs 1,600 cm at (6,2,0,lod3)).
Fixed; tile pixel-identical to CPU, 680/680 + 64/64 triangles, 1,434 GPU-eligible
tiles (was 1,082). Audits: hashes w1/w8, walk 5,642/5,642, thin/protected 0 diffs
vs CPU and WARP, survival 6.15 GB.
Adaptive start: CPU owns the first view while the shader queue is pending.
My runs outside the sandbox (Step17.exe, interleaved):
CPU 1.375 / 1.363 / 1.354 s
GPU 2.040 (first launch of build; 0 GPU tiles, CPU worker sum 6.98 vs 5.05 s)
/ 1.279 / 1.337 / 1.328 s
Cached warm-up queue now 329-632 ms (step 16: ~3 ms), so GPU took 258 tiles
once and 0 twice: cached GPU ~= CPU, a regression from step 16's 1.26 s.
First launch improved 3.09 -> 2.04 s but still loses to CPU. Also adds
STEP16_REPORT.md, missed in 666dfb8.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Timeline showed the post-export window was frame-paced: the 4-batch / 2 ms
per-frame upload budget held CPU-fallback tiles up to 755 ms after they
finished, and 35 GPU batches drained one per frame. Before the first whole
view, GPU work now drains as one batch and CPU fallback starts with GPU input
preparation (--gpu-far-no-cold-drain = paced control).
Luna's race: CPU 1.393 s vs GPU 1.677 s. My interleaved runs (Step15Timeline,
RunGpuFarViewer): GPU 2.902 (first launch) / 1.252 / 1.209 s; CPU 2.196 /
2.250 / 2.061 s. The CPU viewer still uses the paced upload budget at cold
start (stream 674 -> 2250 ms in my runs vs 741 -> 1422 in Luna's), so the
race is not yet fair. Audits: hashes, walk 5,642/5,642, 0 diffs vs CPU/WARP.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
- Stairs: cave, floor, strata and passage relief were effectively summed in overlaps (owner ray
~0.94 voxel of displacement); now partition-of-unity weights: -0.288 voxel there, max detail
gradient 0.323 voxel/voxel. High-detail patch scan 133 -> 0; exact midpoint edge roots kept
(removing them: 7,993 bad roots, 46 patches).
- Fins: exposed skirt triangle scan 1,776 -> 0; thin sheets 0; coarse off-surface panels 0.
- Path hash both directions equal; W1/W8 equal; routes 5642/5642; passage detail ratio 0.850;
far survival passes, bad-shell fails; 30-crossing flight: nearest tile 15.1 ms median, 124 ms max.
Caveats seen in the captures: the owner-ray wall has visibly less relief than before; the fin
before/after captures do not show the fins (evidence is the scan only). Cold StreamFullClipmap
0.758 s vs 0.741 s (+17 ms, Luna's interleaved run). Owner to judge: bin\VoxelForgeLookCheck4.exe.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Not verified as a whole. In the tree when Luna's session ended ("model at capacity"):
- fins: LOD seam skirts could hang into tunnel air; skirt triangles are now clipped against the solid
half-space and carry a per-face mask so a host can hide skirts between same-LOD neighbours.
- choppy LOD0 stairs: marching cubes' linear edge interpolation misplaced vertices where the relief
bends within one cell; edge crossings corrected with the exact field (82 patches, midpoint-edge
errors 18,818 -> 3,288); normal halo now ignores the +-4.125 fallback samples.
- an explicit unclamped Core distance query for audits.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
- The copied generator, mesher and Compat stand-ins compile inside vfcore::detail.
- CheckCoreSymbols.ps1 (run by Build.ps1) fails on any unqualified Unreal-style symbol in the
library: 16,255 on 557f6aa, 0 now. A UE 5.7 CoreMinimal.h + VoxelForgeCore.h smoke compile passes.
- Output identical: 11 hash rows at W1/W8, routes 5642/5642, look/passage/far audits pass,
bad-shell control fails, six captures byte-identical.
Cost, measured by me interleaved (8 workers, cold, 4 runs each, first run excluded): 0.963 s vs
0.937 s for 557f6aa, ~27 ms in shape compilation; same math, so codegen/inlining. To be found.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
- Standalone/Core: public header VoxelForgeCore.h (no Unreal types), static library target,
ResolvedWorldSnapshot (from the schema-v1 JSON), World (compile, generateTile, density/sign,
point/segment solid, classifyBox), host-supplied JobScheduler (the core starts no threads),
sequenced, idempotent EditTransaction returning affected tiles per LOD.
- The viewer, benchmarks, edits and audits run on Core; StreamingWorld stays host code.
- Same as round 11: 11 hash rows at W1 and W8, routes 5642/5642, look/passage/far audits pass,
bad-shell control fails, six fixed captures byte-identical; cold W8 1.025 s vs 1.041 s reference.
- SHARED_CPP_LIBRARY_PLAN.md rewritten for a UBT module compiling the same sources via shims.
Open issue before Unreal: Core's implementation still compiles the Compat stand-ins named like
Unreal types (TArray, FVector...); inside a UE binary that is an ODR clash. Must be namespaced.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
- Far repairs are source-aware (only where the sampled mesh lost a shell) and redundant radius probes
trimmed; ~175k two-triangle repair panels remain (+18 MiB over the no-shell 92.6 MiB).
- Pipelined start builds the shape world once; the 27-tile neighbourhood is ready ~0.73 s from
process start; pipelined whole view is no slower than direct.
- Far audit passes, bad-shell control still fails (16 missing / 16 holes); look audits pass; routes
5642/5642; 60 W1/W8 + direct/pipeline hash fields equal; six fixed captures pixel-identical.
Verified by me head to head (8 workers, cold, BelowNormal 0xFF00, quiet machine, two runs each):
1.40 s vs 1.80 s for the previous default.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
- FarSurvivalAudit: checks every stratum top/bottom boundary (4x4 columns per far tile) and every
in-view far connector (sections <=128 voxels, 4 radial rays) against the emitted mesh. Falsified:
removing one tile's shell triangles -> 16 missing, 16 holes, exit 14. Passes at LOD6/7/8.
- Far connectors follow source radial crossings; strata patches emitted at LOD7/8. Default LOD8.
- Viewer: separate near ShapeWorld for the 27 player tiles, full world built on its own thread;
27/27 tiles hash-identical to the direct build. W1/W8 60 hash fields equal; routes 5642/5642.
Cost, measured by me head to head (8 workers, cold, BelowNormal 0xFF00): LOD8 without the shells
1.17-1.25 s, LOD8 with the shells 1.80 s; payload 92.6 -> 276.7 MiB. The shells are far too
expensive, and the pipelined benchmark whole view is 2.92 s (shape compiled twice): both to fix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
- Per-vertex biome lookup was the payload cost (~365 ms); far tiles now take the biome palette from a
per-tile grid. Same channels (palette, slope, border blend, UV) and ceiling/ground sections.
- Opt-in --max-clip-level=8: 6,498 -> 2,226 tiles, stream 0.876 -> 0.428 s, 1.451 s process wall.
Kept opt-in: six fixed views cannot prove every distant thin stratum survives.
- W1/W8 hashes equal across 11 outputs; routes 5642/5642; five captures pixel-identical, passage
382 px differ (max delta 16).
- Rejected with measurements: LOD6 empty-block proof (-21% samples, no wall gain), far supertiles
(no complete groups exist).
Remaining long pole: shape compile 0.854 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Takeover after the PC reset. Verified at 1 and 8 workers: density, geometry, full payload and
topology hashes equal for all nine outputs; routes 5642/5642; crater transaction 17 ms; player
neighbourhood mesh-ready 23 ms after streaming starts.
- The plugin's materials consume UV, the packed biome/slope/blend colours and the ceiling/ground
sections, so the viewer-only skips stay off by default (diagnostic only: 1.53 s without them).
- UVs staged as FVector2f (matches the game stream; -20 MB, no speed change). Far-bin reserve 24x.
- LOD6 is the long pole: 32.7 M density samples, 1.83 M vertices, ~1.1 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Verified before the reset (density + full mesh hashes equal at 1/8/16 workers, routes 5642/5642,
look audits green):
- merged field lookups, parallel mask merge, compact inline spatial bins, line-local relief hash
cache, lattice-specific mesher entry, fused sign+magnitude evaluation, far connectors as one
capsule each (181,836 -> 16,491 far primitives).
- crater edit (sphere+box+capsule, one transaction across LOD0/1): 15-27 ms; player 3x3x3
neighbourhood meshes ready ~18 ms after streaming starts.
NOT yet verified (in the tree, unhashed): standalone-only skips of UV/colour arrays and of the
ceiling/ground material grouping (best cold whole view 0.958 s @16 workers) and a far-bin reserve
tweak. Those skips drop data the plugin's materials may need: not a game-path win as-is.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
- Removed the thin automatic room-support slabs that left floating caps (in-view detached thin
sheets 9 -> 0; the audit now also fails on isolated small components).
- Bounded 16-voxel floor/wall blend (18 rejected: it created a sub-voxel sheet), 3-voxel
strata/cave blend, C1 tunnel radius profile (0 radius/slope steps at joins).
- Surface relief: broad octave with amplitude modulation instead of uniform bumps.
- Cold, no cache: 64 m full detail 0.924 s (8 workers) / 0.715 s (16); whole view 2.016 / 1.605 s.
- 1-worker and 6-worker hashes match; 5,642/5,642 passage samples open.
Not solved: regular bands in the narrow passage shafts, a few slivers on tile boundaries, faint
wall/floor plates; close-view mean surface distance to the owner field got worse (see REPORT.md).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
ClipRadius override 3 (64 m at 25 cm). Uncached full detail 20.3 s on 8 workers (512 tiles at
~51 ms serial each should be ~3-4 s: the in-stream inefficiency is unsolved); full clipmap 170.6 s.
The 1 s / 5 s targets were met only from an on-disk mesh cache (.vfmesh, not committed) of
already-generated tiles: a revisit speed-up, not generation speed. Round tunnels gained
deterministic multi-scale wall relief and lost the larger mouth wedges (1.54% of lattice signs
changed, mean 0.20 m). Hashes stable across runs and 1 vs 6 workers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Cause found by profiling: room/SDF graph-cache churn (exact XY bounds + sample order rebuilt it
constantly). Region grouping (one LOD3 tile: 1,084 -> 229 builds), taller-tile traversal for LOD6
(36,480 -> 12,160), extent-dependent cache slots, lambdas instead of recursive std::function.
Serial ms/tile, 64 nearest tiles per level (before -> after):
L0 50.4 -> 50.6, L1 112 -> 110, L2 24 -> 22, L3 223 -> 73, L4 269 -> 134, L5 393 -> 127, L6 283 -> 145.
Hashes of all levels unchanged; parity 3,436/3,436 bit-identical.
Full clipmap, 16 workers BelowNormal: full detail 66.6 s, all 12,054 tiles 175 s. In-stream L0 tail is
high (median 83 ms, mean 198 ms, p95 669 ms) vs 51 ms serial.
Streaming: recentre keeps the intersection, generates only missing tiles nearest/finest first;
one crossing = exactly 441 new tiles in 6.1 s, no stale results.
Lighting: normals already come from the density gradient; the striping was the squared headlamp;
broad linear fill + warmer palette. Daytime CPU rule documented in Standalone/README.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Hashed TMap/TSet and real TInlineAllocator storage did not move level 0 (72.1 -> 74.1 s, within
noise): the slowness is elsewhere. StreamingWorld.h started (first build hit an int64/uint64
mismatch), library plan drafted (SHARED_CPP_LIBRARY_PLAN.md), before-normal captures saved.
Uncompiled/unmeasured parts: streaming. Resume by forking session 01a0fd2e-3c3b-7050-b4e9-628fdcc4d553.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Fr0zkaandClaude Opus 5.5