- Headless and Unreal schedulers drain the collision-critical ring, then LOD0-2, then far LODs
at low priority. LOD0-2 ready (W8 medians): SW 4.36->1.76, FI 6.89->5.26, TN 4.88->3.22,
VS 2.71->1.56 s. Whole view roughly unchanged (FI/VS slightly later).
KNOWN RISK: far tiles wait until every near tile is ready, so far LODs may starve while
moving fast (to address next).
- Exact composer trims: tile-level upper-mouth candidate list, segment-index min/max rejection
before lower_bound, 64-entry thread-local value-noise corner cache. Passage share 65 -> 63%.
- World exact: 24/24 hash matrix + pinned L75 owner hash. Fast tier PASS. Linux UE host builds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
VoxelHash::Cell/Pair, legacy Cellular3D and the shared Core cellular hash added 0x7FFFFFFF to a signed
int32 before casting to uint32; overflow is undefined, so Clang and GCC placed cave passage 1 differently
(first differing sample at LOD1 tile (0,64,-512)). The bias is now done in unsigned arithmetic, in both the
standalone copy and the Unreal plugin header (Source/VoxelForge/Public/VoxelCaveMorphology.h). Native
caves, motifs and lattice now stream LOD0-8 with skirts and cross-level transitions. New verify.py
--speed-matrix runs every generator on one asserted-identical view (5 interleaved runs): motifs 1.33,
profiles 1.30-1.32, caves 1.41, natural SW 1.67, lattice 1.70, FI 4.93 s whole view. fast/medium/full PASS,
Clang/GCC peer hashes equal. Research R29 (open-edge ledge) and R30 (compiler-divergence suspects).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
The shared passage carve took the minimum of piecewise-linear capsules (scallops at joins) and added
wall relief on regular 8 m / 4 m octaves, which read as evenly spaced ribs. Interior sweep joins are
now Catmull-Rom smoothed and the relief uses irregular rotated 43/29/19 m scales; the natural wobble
remains. New axial ripple audit (0.125 m radius samples, 0.5-8 m periods): beaded control and L34
geometry fail, all 24 passage rows pass. FI mouth tread apron tapered. 30/30 determinism, 16/16
invariants, ring audits LOD0-8 pass; cost within noise.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSFYgsdW8ejZSyiMf4qRSX
Cold first use: near base 4.2 ms, near detail 1,595 ms (the floor), connector
rings 630 -> 409, skirt classify+emit ~1,409 -> 366. Production warm-up is
still ~2.03-2.06 s; whole view median CPU 1.356 s vs hybrid 3.044 s.
Near 960 + far 1,114 tiles: 0 topology/protected/thin diffs vs CPU and WARP;
hashes, walk 5,642/5,642, survival pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Near LOD0/1 (detail band, relief, root refinement, normals) now evaluated and
meshed on the GPU. 960 near tiles: zero sign/topology differences vs CPU and
WARP, vertex max 0.0049 mm (was 15 cm); far survival audit passes; CPU hashes
44/44 unchanged. Far connector-panel outlier (64.56 cm, LOD8) remains a
tessellation difference, and remote captures show missing far geometry on GPU.
Cold readiness (viewer clock, my interleaved runs): CPU 2.51/2.49/2.40 s,
GPU 5.05/2.68/2.80 s. The ~2.1 s Present stall hit only the first GPU launch.
Viewer spends ~1.13 s before shape compile starts (startup checks), which is
the gap between the 0.95 s benchmark and ~2.4 s on-screen.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Seven LOD2-8 tiles: RX 7900 XTX field+meshing 0.335 ms vs CPU generateTile 15.81 ms; repeat runs
bit-identical; CPU W1/W8 44/44 hashes unchanged. Not yet a real gain: rock/air signs still come from
CPU masks for every sample. Viewer does not draw GPU tiles yet; no whole-view timing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
- Compute shaders classify cells from CPU sign bits, scan output offsets, emit vertex/index buffers
and indirect-draw args. Two 32^3 tiles: 0 topology differences hardware vs WARP; two runs on the
RX 7900 XTX identical; max CPU/GPU vertex distance 0.000122 cm (LOD0), 0.000244 cm (LOD1).
- GPU kernels 0.049 / 0.071 ms vs CPU tile 2.33 / 1.99 ms, but CPU input preparation dominates
(7.25 / 2.84 ms wall): no whole-view gain until the field itself runs on the GPU.
- VoxelForgeFieldShared.h: first shared C++/HLSL field math (integer noise hash, ...).
CPU path unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Fr0zkaandClaude Opus 5.5