Step-2 (LOD1) tiles of native SurfaceWorld now prove 8-cell blocks uniform from exact heightfield
column extrema over the padded block (cells + normal halo) and skip their density fill; features,
edits, mixed strata and Unknown verdicts still fill. LOD1 samples 42,875 -> 2,063 per tile.
Audit: 80,502 skipped blocks (3 seeds) + 9,216 island blocks force-filled, 0 violations. All
SurfaceWorld and FloatingIslands hashes unchanged at W1/W8/W12, Clang off and GCC default.
Reviewer interleaved W12, no hashing: whole view 1.439 -> 1.409 s. L8 phase probes removed;
--profile-lod-work and --audit-uniform-block-proofs are opt-in.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSFYgsdW8ejZSyiMf4qRSX
SharedColumnCache.h: a world-owned, bounded 4-way second-tier cache of exact column values
(heightfield and floating islands), keyed by recipe/plan identity plus content hash, XY and island
id. Lock-free seqlock reads; set-local spin lock on misses. The FLazy thread-local tier stays
first. 83.6% of repeated column work was cross-worker; formula evaluations -60.9%.
W12 L1 view: Luna 2.165 -> 1.754 s; reviewer interleaved 2.059 -> 1.722 s. 8 m crossing new tiles
82 -> 52 ms. +66 MiB peak RSS. All hashes (whole view, per-LOD, exact queries, islands seed 0)
identical at W1/W8/W12, Clang off and GCC default. MSVC/UBT not compiled here.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSFYgsdW8ejZSyiMf4qRSX
Linux offscreen capture (EGL/GLES pbuffer; llvmpipe in the sandbox) and an opt-in screen-space gate
for LOD>=3 (normal distance projected at each LOD's closest display distance, 1 px at 1440p =
0.61 mrad). The gate rejects the S13 octave-drop and stride-2 controls, fault injections, and a
vertical-vs-screen control. Finding: today's far meshes differ from the field by up to 92 px, and
LOD swaps snap mesa rims by tens of px (p50 0 px), a silhouette effect of plain LOD switching. The
octave cutoff was dropped: nothing qualified, and octaves are a small share of far time. Off-path
cost: reviewer interleaved W12 within noise vs L4 (+0.6% median). Hashes = L4.
Largest far lever measured: column work outside noise, 34% of far worker time (duplicate XY across tiles).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ED33VnqepD4YY6gDAAA832
VoxelForgeCoreThreadLocal::FLazy<T> (new public header ThreadLocalStorage.h) makes every large
thread_local cache a pointer that heap-allocates on that thread's first use (still worker-private,
lock-free). Static TLS 17.33 MiB -> 43 KiB (initialised image 5.8 MiB -> 21 KiB); worker start
5.2 -> 0.12 ms; idle worker RSS 17.7 MiB -> 52 KiB. Linux workers back to 8 MiB stacks.
World unchanged (L1 hashes Clang off/GCC default, W1/W12; FloatingIslands seed 0).
Reviewer interleaved W12: 2.256 -> 2.091 s. MSVC/UBT not compiled here.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ED33VnqepD4YY6gDAAA832
CoreNamespacePreamble.h applies per-compiler contraction-off pragmas (Clang, MSVC, GCC; unknown
compiler = #error) and a fused-multiply-add self-test that World::compile runs before building.
Clang -ffp-contract=on and GCC's default (fast) now reproduce the L1/Windows hashes (they diverged
in L2); Clang -ffp-contract=fast ignores the pragma and is refused at startup (negative control).
Seed sweep with every cube on real terrain: 48 seeds x 2 recipes x 3 builds x W1/W12, 576 runs,
0 mismatches, 0 empty meshes. Cost +0.5% (2.326 -> 2.337 s, interleaved). World unchanged.
MSVC/UBT not compiled here: Windows to-do list in Linux3/REPORT.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clang -ffp-contract=on/fast and GCC's default (fast) all change the density hash on the L1 view;
contraction off reproduces L1 (= Windows S13) at W1/W12. Seed sweep: 48 FloatingIslands seeds and
14 effective SurfaceWorld seeds (34 of 48 fell on a seed-independent cube) match across Clang/GCC
and W1/W12. Standalone gains --seed/--recipe/--render-distance-chunks (viewer + headless) and a
Windows replay script. Linux standalone workers get 32 MiB pthread stacks (TLS is 17.3 MiB/thread),
so no external stack limit is needed. No Core or world-output change; MSVC compile unverified.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Headless Linux benchmark/hash tool (Clang 19 / GCC 14) sharing the viewer's plan builder and
streaming pipeline (moved verbatim into StreamBenchmarkPipeline.h / WorldAssemblyPrototype.h).
Owner Round17 view: density/mesh/payload/topology hashes equal the Windows S13 reference at
W1/W8/W12 (Clang, 3 runs each) and W8 (GCC), with -ffp-contract=off. No output change.
Baseline (i9-12900H, CPUs 0-11, powersave/balance_performance): whole view W8 3.562 s median,
W12 3.512 s, W1 16.555 s; peak RSS < 605 MiB.
Review: the default-stack crash is 17.3 MiB of static TLS per thread (glibc puts it in the
thread stack; HeightfieldColumnCache 6.8 MB), not recursion -- open, after S14.
MSVC compile of the viewer refactor not verified (no MSVC here).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Exclusive per-stage tile timers and probes. Findings: LOD0 is MC-exclusive
(10.4 of 14 ms p95, 243 ns/sample); LOD4's 3x cost is column-cache misses
(8,224 vs ~1,600-1,900 at LOD3/5), cap refine secondary; per-tile feature lists
cost ~0.001 ms; cross-worker column repeats common (277k at LOD0). Global octave
drop and stride-2 interpolation probes fail the 2 cm gate. Worker cap 1 kept
(cap 0 worse 40 m/s tick tail). Report: Standalone/Research/WorldAssemblerM3/Round17.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Planar far horizon samples the LOD8 interior at the MC cell spacing, keeping
128-voxel boundary strips; far recipe v3. LOD8 p95 59.8 -> 13.9 ms, spawn
worker-seconds ~41 -> 26.9, full set 7.8 -> 7.1 s, ring 0.49 s, LOD0 1.43 s,
40 m/s ring 75/75. S3/S4 2 cm agreement, 100% cap coverage, W1/W8/W12 and
W1/W8/W16 determinism pass. Report: Standalone/Research/WorldAssemblerM3/Round14.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Direct per-tile trace on the BuildHost actor (owner plan, 2,578 keys): full set
17.4 s legacy; gated coarse-first (after the 9-tile walkable ring) 11.7 s with
the ring still at 0.63 s, but full LOD0 slips 2.3 -> 10.2 s. Startup apply
boost changed nothing (0.4%). 37.6 worker-s over 17.4 s on 10 slots = ~3.3
busy: the floor is admission/work supply. S5's 251 ms staircase was the trace
sampler. Fixed CVar parsing in headless BeginPlay and the 60 m/s crossing
instrument. All three knobs stay opt-in.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
S3's far cap sat up to 8.2 m above the near roof and read as floating pieces.
The far render cap now projects onto the owning strate's upper boundary
(density unchanged), on a global XY lattice, LOD4 16-voxel grid + LOD5-8
128-voxel. Cap agreement 0.05 cm (100 cm control detected), 2520/2520 upward
rays hit roof on the owner walk. Far recipe v2; LOD0-3 hashes equal S2/S3.
Tile p95 flat; cold clipmap +4.4%. Far-cliff fluting left open (coarse MC step).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Far-shell repair (81-96% of S2's LOD7/8 tile time) is skipped for certified
SurfaceWorld tiles; ground and cap come from the shared planar heightfield.
Near LOD0-3 hashes equal S2; far recipe version 1 salts LOD5+ identities.
Cold first view 6.6 -> 3.4 s (aim 1.57 s). Not a visual sign-off: distant cap
shows floating pieces, sprint streaming arrives late.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
AVoxelWorld::CoreVoxelSizeCm (default 25, 32 voxels/tile) flows through plan, field,
mesher, content, water, atmosphere, edits and collision. Staged-owner BuildHost passes at
20 cm (SurfaceWorld + islands, W1/W8/W16, anchors, BP edits, collision) with a fixed-25cm
negative control; default-size hashes match round 4; Core off stays legacy 12.5 cm.
Reusable BuildWorldAssemblyFeatureTileListV1; wired into FloatingIslands only so far.
256-island stress: tile p95 112.5 -> 2.0 ms, 0 mismatches, dropped-feature control fails.
Content placement measured (no populated cells): island ground search 7.25 ms/cell vs
legacy 0.29 ms -- open. Outside-review evaluation in Round5/REPORT.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Islands read as separate (91.89% open sky at mid-height), underside cavity closed,
rim teeth were capture shading. Island manifest import + BuildHost proof passed on the
pre-cache tree; M1 SurfaceWorld and M2 V1/V2 gates passed.
Known open: final XY-column cache compiled in the owner build but is unverified;
island LOD0 p95 ~155 ms misses the 133 ms sprint deadline (legacy ~43 ms); the island
manifest only fits the BuildHost fixture layout (owner's world: Core init fails,
empty world) -> round 4 builds plans inside Unreal.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
The round-3 'field contact error' measured hit-to-anchor distance; V2 contacts are 0 cm and on the zero
isosurface. Tagged-tile cost 5.7-6.7x -> 1.175-1.270x MC; organic 0.998x. Bound, shelf-thickness and route
audits rerun with negative controls; W1/W8/W16 stable; V1 and MarchingCubesV1 unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Owner saw content floating above native terrain, spiky triangles, reversed carve/fill and late relief.
Causes: shared mesher placed Core tiles on a 12.5 cm grid; content/atmosphere asked strate ranges in
12.5 cm voxels; Core edits add density where the legacy diff subtracts. Horizon LOD8 bridge edge
5.657 -> 2.265 cells. Relief mesh p95 60.8 -> 42.0 ms (LOD0). New BuildHost checks for scale, height,
blueprint carve/fill and per-kind streaming lag.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna, M1 round 4 (resumed after the night stop):
- Cliff: sub-voxel audit (29,929 half-voxel XY columns, min run 3.75 voxels, 0 subvoxel runs;
old taper control exposes 20,963) and a root-line shelf at the two-voxel minimum.
- Horizon slab = aliased LOD8 cap ceiling. The first fix hid the ceiling at LOD8 (challenged:
enclosed, interior-lit strates); now the LOD8 cap is rebuilt deterministically from the
native heightfield, skirts preserved; looking up and out shows the cap present.
- Routes: review manifest with 33 stable-ID route points + capsule/support constraints; capsule
sweep 0 blocked samples / 0 unsupported centres, negative control detected.
- Server path: World::densityAt answers without render meshes. Physics cooking not available
standalone.
- Hashes (evaluator + 2,226-tile stream) identical at W1/W8/W16. LOD8: 4.446 s for 529 tiles,
14.25 M field calls, 165 unknown tiles.
- Captures at 60 fps cap + vsync (owner's GPU coil whine). One cliff probe crossed the ~8 GB
guard (7.69 GiB) and was stopped before capturing.
My review: cap view matches legacy closely, no slab; native cliff is a clean attached ledge
(cleaner than legacy); a few small spikes on the ridge top.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna, M1 repair round: the rectangular raised block and the broken cliff shared one cause,
a type-2 solid AABB in the generated plan; it is now a tapered, terrain-following shelf, and
a negative-control audit catches the old box behaviour; seam classification check passes.
Bound audit 298 boxes / 8,046 samples, 0 misses. Evaluator and 120-tile density/mesh/payload/
topology hashes identical at W1/W8/W16. Cave regression: walkability 5,642/5,642, 64/64 tiles.
Legacy mismatch unchanged (0 biome mismatches, 18.14% signs).
Speed (matched workloads): 120-tile full detail 0.089 s native vs 0.625 s legacy; 529-tile
clipmap 3.858 s vs 11.687 s -- 2.29 s above the ~1.57 s aim; LOD8 dominates (165 unknown
tiles, 13.27 M field calls, 10.88 M column-cache misses).
My capture review: coast matches legacy closely; surface clean but a dark flat slab on the far
horizon; cliff shelf renders as a stack of thin plates with back faces showing (sub-voxel
shelf thickness suspected).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna, M1 round 1 of 4 (plan in Standalone/ASSEMBLER_M1_REPORT.md):
- Versioned fixed-scale world manifest (the shipped plan format) and the Core block
contract: point evaluation, interval-or-unknown, influence bound, detail envelope, channels.
- Interval audit: 256 boxes / 6,912 samples, 0 missed, 0 unknown; a deliberately false bound
is detected (negative control).
- Native Heightfield2D with biome input: 4,096-point hash fae2721c052aec5c identical at
W1/W8/W16; point queries 2.059 ms vs 62.115 ms legacy -- but NOT a match yet (mean density
error 274 voxels, 2,261 sign mismatches vs the legacy surface). 45 mesh-free Core queries
match direct evaluation exactly. Existing cave world unchanged (walkability, 64/64 tiles).
- Open: collision readiness, a valid full-view timing (viewer captures were occluded/empty;
top-down height maps are the only usable pair).
Also carries the design doc's owner decisions (fixed water; player edits unrestricted, route
law for automatic changes only) and three plugin files + BuildHost scripts that earlier round
commits had missed (also fixed on unreal/core-integration as 2574b56).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna, round 5 goal 1 only (stopped by owner decision before the SurfaceWorld port):
- Worm and Cascading passages now pass walkability; the wide floor stays clipped while the
narrow route core stays open.
- Hidden third failure found and fixed: long VerticalShafts diagonal legs were dropped from
Core's spatial bins by a 100,000-bin safety cutoff (invisible to density queries);
conservative bins restore them (18,570/18,570).
- All 15 saved probes pass a walkability-only gate; build-time walkability gate added.
NOT re-verified before the stop: W1/W8/W16 hashes, BuildHost UBT/actor run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna: the viewer reset itself to BelowNormal (since step 3) and 0xFF00 (since step 11)
at startup, overriding the runner; every earlier "full machine" race started on 8 CPUs
until the runner reapplied the mask. Fixed in viewer and agreement runner. Core
worker clamp raised to 16 (W1/W8/W16 hashes identical). Shader binaries now pinned per
executable (precompiled/<exe>/); Step27 control rebuilt from 7d1808f with only the
bundle lookup backported. Tried CPU fallback for skirted near tiles: -0.74 GiB GPU
memory but 2,934-4,086 missing pixels in four remote views, so GPU skirts stay default.
Audits: hashes W1/W8/W16, walk 5,642/5,642, near/far HW+WARP 0 diffs, both survivals
pass (7.00 GiB), ten captures 0 missing pixels.
My runs outside the sandbox (driver cache warm, marker removed first):
0xFFFF Normal: CPU 1.121 / 1.081 / 1.040 s
GPU miss 1.324 s; cached 1.340 / 1.353 / 1.293 s (2,073-2,074 GPU tiles,
timestamps 463-471 ms)
0xFF00 BelowNormal: CPU 1.305 / 1.302 s; GPU cached 1.678 / 1.692 / 1.756 s
Same-session, own shader bundles, 0xFFFF: Step27 GPU cached 1.205 / 1.212 / 1.198 / 1.180 s;
Step29 GPU cached 1.359 / 1.367 / 1.366 / 1.333 s
=> the GPU route now loses to the CPU route by ~0.25 s on the full machine.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna: near skirts were CPU-only by implementation gap, not correctness. Ported with
near-resolution refinement and overflow fallback; near-probe preparation split out so
the skirt kernels no longer inline the near field (first attempt: two fxc /O3 compiles
ran 10.5 h and were stopped). Shape compile proven identical at W1/W8/W16.
-SkipCpuDetailBand retired (154/640 skirt audit failures, +174 ms planning).
Audits: hashes W1/W8/W16, walk 5,642/5,642, whole-near 960 tiles + 640 skirt faces
0 sign/topology diffs (one 0.001 cm vertex-key delta), far/thin 0 diffs, both
survivals pass with 2,074 candidates at 7.00 / 7.02 GiB (was ~6.2), ten captures pass.
My runs outside the sandbox (Step28.exe). NOTE the machine was ~25% slower than last
night for every build (step 27 CPU full machine 0.97-1.02 s at night, 1.22-1.29 s now):
0xFF00: CPU 1.337 / 1.291 / 1.295 s; GPU miss 1.337 s;
GPU cached 1.663 / 1.424 / 1.617 s (GPU tiles 1,427 / 2,072 / 1,222)
0xFFFF: CPU 1.221 / 1.231 / 1.236 s (step 27 same session 1.285 / 1.222 s)
GPU cached 1.582 / 1.550 / 1.416, 1.635 / 1.441 / 1.470 / 1.445 s
(GPU tiles vary 593-2,074; timestamps up to ~470 ms)
=> GPU route now slower than the CPU route measured alongside it. Step 27 GPU control
was not clean (both exes load the same precompiled .cso files).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna: normal completion path reads the raw vertex count before residency; an
over-cap count or failed readback falls back to CPU meshing. Max seen 18,153 of
19,456 vertices at (2,-9,-4,lod1) -- 1,303 headroom for this seed only.
Consumer map: 320 GPU near, 640 CPU near (skirted), 1,114 GPU far, 152 CPU far.
Exceptions are exported separately from roughness. Opt-in -SkipCpuDetailBand skips
CPU roughness for GPU-owned near visuals (lazy exact reconstruction on query).
Runner gains -AffinityMask (default 0xFF00) and -NormalPriority.
My runs outside the sandbox (Step27.exe, marker removed first):
half machine (0xFF00, BelowNormal), interleaved:
CPU 1.374 / 1.285 / 1.296 s
GPU miss 1.330 s; cached 1.151 / 1.143 / 1.129 s (1,434 GPU tiles)
GPU cached + -SkipCpuDetailBand 1.316 / 1.303 / 1.307 s -- SLOWER: detail apply
331 -> 297 ms, but intervals+signs 112 -> 286 ms (the plan), shape 557 -> 692 ms
full machine (0xFFFF, Normal), interleaved:
CPU 1.004 / 1.015 / 0.974 s
GPU cached 0.931 / 0.930 / 0.925 s -- first sub-second launches
Audits (Luna): hashes w1/w8, walk 5,642/5,642, near (skip) / far 0 sign/topology
diffs vs CPU and WARP, auto + cold survival pass, ten captures pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna: on a marker miss the deferred warmup no longer gates the initial view
(starts after the full CPU view is presented). The forced-cold DXGI 0x887A0005 was
ours: GetDeviceRemovedReason = DEVICE_HUNG after the audit readback of tile
(-7,-7,0,lod8), slot 1433 -- the readback copied the theoretical 61,440-vertex maximum
from a 19,456-vertex pooled slot, 41,984 vertices past the end of the last chunk's
buffer. Readback now clamps to slot capacity (audit-only path; normal launches never
read back). Auto and forced-cold survival both pass (1,434 submissions, ~44.6 s, <8 GB).
Added a no-write option for the warmup marker. Direction answer in LUNA_DIALOGUE.md.
My runs outside the sandbox (Step26.exe, marker removed first, interleaved, -Timeline):
CPU 1.299 / 1.291 / 1.266 s
GPU marker miss 1.319 s (all-CPU, = CPU again; step 25 was 2.015 s)
GPU cached 1.151 / 1.122 / 1.138 s, all 1,434 GPU tiles (timestamps 243 / 307 / 241 ms,
max dispatch 6.3 / 66.7 / 5.2 ms)
Audits: hashes w1/w8, walk 5,642/5,642, near/far 0 sign/topology diffs vs CPU and
WARP on protected/thin, ten captures pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna: shape_compile profiled (8 workers, detail band ~323 ms of ~545 ms over 9.39M
candidates; roughness dominates). Exact vertical roughness endpoint reuse in
ShapeWorldInternal.h, hashes unchanged: ~548-577 -> ~535-548 ms. Overlapping
pipeline/pool with shape or streaming made things slower, so it stays serial (pool
precompile now 4-20 ms). Export starts before the first CPU frame. Marker-miss stall
reproduced and gated differently; the forced-cold survival audit now hits
DXGI 0x887A0005 (device removed) on a timestamp query and is NOT passing. Auto
survival passes (6.16 GB). The vertex count at InitialClipmapReady counts CPU
batches only (GPU-resident tiles excluded) -- confirmed counting, not missing geometry.
My runs outside the sandbox (Step25.exe, marker removed first, interleaved, -Timeline):
CPU 1.329 / 1.295 / 1.318 s
GPU cached 1.152 / 1.170 / 1.137 s, all 1,434 GPU tiles (timestamps 292 / 260 / 356 ms,
max dispatch 19.0 / 15.9 / 75.2 ms)
GPU marker miss 2.015 s -- REGRESSION vs step 24's 1.336 s. CPU had all 2,226 tiles by
~1.3 s, but near_all_tiles_ready waited for the post-view shader warmup
(queue 1,025 ms, finished 1,955 ms; handoff cpu_started=1434 takeover=0).
Audits: hashes w1/w8, walk 5,642/5,642, near 960 / far 1,114 0 diffs vs CPU and
WARP, thin 353, ten captures 2+2 edge pixels as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna: warmup policy (marker read, hashes, adapter query) now resolves on a
background thread; startup fflush(stdout) removed (a redirected flush measured
136 ms). Device-init -> generation-setup gap 0.056 ms on the miss route.
Deferred-batch replay runner reproduces 85-86% of the viewer's GPU time
(240 vs 279 ms); the old "10x" compared one LOD2 tile to a mixed LOD0-8 mean.
Remaining sequence-dependent spikes (11-29 ms on tiles that take ~0.1 ms alone)
are unexplained. Open: a marker-miss survival audit once stalled after
[CpuReadyFrameBeforeGpu].
My runs outside the sandbox (Step24.exe, marker removed first, interleaved, -Timeline):
CPU 1.350 / 1.281 / 1.266 s
GPU marker miss 1.336 s (all-CPU fallback, generation setup at 88 ms)
GPU cached 1.167 / 1.166 / 1.168 s (GPU timestamps 220 / 255 / 205 ms,
max dispatch 5.6 / 22.0 / 1.0 ms; 1,134-1,236 GPU tiles)
Steadier than step 23 (1.07-1.17) but not faster; <=1.0 s not reached.
Audits: hashes w1/w8, walk 5,642/5,642, near 960 / far 1,114 0 diffs vs CPU and
WARP, thin 353, survival 6.19 GB, ten captures 2+2 edge pixels as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Connector-ring kernel skipped when a tile has no rings, no connector vertices and no
visible skirts (skirted tiles keep it: an earlier variant lost remote coverage, caught
by the captures). Outlier controls: the step 22 67 ms tile runs in ~23 us isolated,
also after a pool touch; no-present control did not remove outliers. Cause unknown.
My runs outside the sandbox (Step23.exe, marker removed first, interleaved, -Timeline):
CPU 1.360 / 1.294 / 1.287 s
GPU cached 1.168 / 1.072 / 1.151 s (GPU timestamps 295 / 245 / 264 ms,
max dispatch 52.7 / 9.2 / 35.0 ms)
GPU marker miss 1.449 s. Measured: on the miss route, device-init end to
generation-setup begin takes ~132-155 ms (s22 a/e, s23 a) vs ~3.5 ms on the
match route; s22 f (miss) was 3.5 ms. That gap explains the slow first launches.
Audits: hashes w1/w8, walk 5,642/5,642, near 960 / far 1,114 0 diffs vs CPU and
WARP, thin 353, survival ~6.14 GB, ten captures as step 22.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Per-batch GPU timeline (submit, timestamps, query, consume). Startup Draws submit
several ready batches per pump (80 ms startup budget, 16 ms per batch); the output
pool is reserved before the warm-up-pending deferral.
My runs outside the sandbox (Step22.exe, marker removed first, interleaved):
CPU 1.295 / 1.314 / 1.284 / 1.342 s
GPU cached (marker match, -Timeline): 1.197 / 1.121 / 1.133 s
(1,124 / 1,380 / 1,415 GPU tiles; output pool at 655-680 ms in all three)
GPU marker miss: 1.446 (-Timeline), 1.447, 1.312 s (no -Timeline). The 1.447 run's
shape compile ended at 776 ms vs 634-643 ms in its neighbours, so this is
startup variance, not a clear regression; to be watched.
Known tradeoff: startup pumps can stall the main thread 65-108 ms (submit calls run
past the budget); fine before the first view, to check in an interactive viewer.
Audits: hashes w1/w8, walk 5,642/5,642, near 960 / far 1,114 0 diffs vs CPU and
WARP, thin 353, survival ~6.2 GiB, ten captures as step 18.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
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
ApplyDetailBand claims eight adjacent columns per batch, keeping each column's scalar
order: shape stage ~586 -> ~566 ms in Luna's profile. Hashes w1/w8 identical.
My runs outside the sandbox (Step20.exe, marker removed first, interleaved):
CPU 1.319 / 1.287 / 1.288 s
GPU 1.320 (marker miss, 0 GPU tiles) / 1.244 / 1.215 / 1.219 s (match, ~1,400 GPU)
Shape compile 549-570 ms in both modes. The 669-821 ms gap in cached runs is still
there: export starts only after [ChunkCrossingWork] reports the ~840 CPU tiles
(LOD0-2) finished, so GPU export waits behind CPU meshing on the shared workers.
Slowest GPU tile 1.7 / 18.4 / 27.3 ms in runs b/c/d.
Audits: hashes, walk 5,642/5,642, near 960 / far 1,114 0 diffs vs CPU and WARP,
thin 353, survival 6.14 GB, ten captures as step 18. Far vertex-distance diagnostic:
247 bound failures (max 64.56 cm), unchanged since step 16.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
GPU tile inputs are published as workers finish (LOD then camera distance); first
batch at 32 ready, later batches up to 256. Export wall 224 -> ~105 ms. A camera-near
partial shape compile was tried and removed (duplicated work, missing inputs).
My runs outside the sandbox (Step19Stream.exe, marker removed first, interleaved):
CPU 1.380 / 1.332 / 1.302 s
GPU 1.358 (marker miss: 0 GPU tiles) / 1.409 / 1.263 / 1.246 s (marker match,
~1,400 GPU tiles, warm-up queue 5-6 ms)
Cached timeline (run d): shape compile 70-652 ms, idle gap 698-850 ms, export
850-954 ms, GPU tail to 1,246 ms (235 ms GPU timestamps). Run b: one 15 ms tile,
ready 1.409 s. Target <= 1.0 s not met.
Audits: hashes w1/w8, walk 5,642/5,642, near 960 / far 1,114 0 diffs vs CPU and
WARP, thin 353, survival pass, ten captures as step 18, peak 6.17 GB.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
App-owned marker (bin/VoxelForgeGpuWarmup.state: adapter, driver, shader sources and
binaries). Miss -> CPU builds the first view, shader warm-up runs after it. Match ->
warm-up overlaps startup, at Normal thread priority (was Idle).
My runs outside the sandbox (Step18.exe, marker removed first, interleaved):
CPU 1.395 / 1.363 / 1.329 s
GPU 1.409 (marker miss, cold driver: 0 GPU tiles, warm-up 683 ms after present)
/ 1.290 / 1.278 / 1.289 s (marker match: warm-up queue 4.6 / 3.6 / 191 ms,
GPU took 1,394 / 1,410 / 1,115 tiles, CPU mesh sum 5.1 -> 1.3-2.0 s,
GPU busy 206-227 ms)
Cached target was <= 1.26 s: not met. With the GPU taking almost every tile, the
wall barely moves, so the rest of the 1.28 s is outside meshing. Known risk: if the
marker matches but the driver cache was evicted, the launch pays ~2.2 s (Luna's
sandbox runs).
Audits: hashes w1/w8, walk 5,642/5,642, near 960 / far 1,114 tiles 0 diffs vs CPU
and WARP, thin 353 cells, survival 6.14 GB, ten captures (2+2 edge pixels as step 17).
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
CPU uploads now bypass the 4-batch / 2 ms cap until the first resident view.
Luna's race (inside the Codex sandbox): CPU 1.354 s vs GPU 3.152 s, warm-up
queue 1.16-1.21 s on every launch. My runs outside the sandbox, same binary
(Step16FairRace, RunGpuFarViewer, interleaved): CPU 1.497 / 1.359 / 1.334 s;
GPU 3.088 (first launch of this build, warm-up 1,287 ms) / 1.262 / 1.260 s
(warm-up 3.2 / 3.1 ms). AMD DxCache was written at my GPU launches (08:46:57,
09:33:37) and never during Luna's runs: the sandbox cannot persist the driver
shader cache, which is why every one of Luna's launches paid the compile.
Audits: hashes, walk 5,642/5,642, 0 diffs vs CPU/WARP, survival 5.10 GB.
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
Per-kernel fences in the production viewer put 1.73-1.98 s on the near-detail
Flush (shader creation 17-19 ms, dispatch calls ~0): skipping only that
dispatch cut idle warm-up 1.893 -> 0.521 s. Near detail rewritten with a
shared owner-roughness loop and shared Perlin samples: DXBC 448,588 -> 120,860
bytes, cold first use 1,596 -> 709 ms. Warm-up median 1.052 s, yet whole view
is unchanged (CPU 1.352 s vs hybrid 3.196 s): after input export the hybrid
spends 1.78-2.00 s streaming 1,082 GPU + 1,144 CPU-fallback tiles.
Audits: 0 diffs near/far vs CPU and WARP, hashes, walk 5,642/5,642, survival.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Split ordinary far evaluation from near-detail evaluation; [loop] in connector
refinement and skirt clipping. Cold first use: EvaluateFarField 1704 -> 3.3 ms,
connector rings 775 -> 630, skirt emit 948 -> 909, near detail 1783 ms (the
floor). Warm-up still ~2.2 s; whole view CPU 1.351 s vs hybrid 3.210 s.
Output-slot reuse fixed: crossing probe 1/1, 2,226/2,226 tiles. Audits: 0
topology/protected/thin diffs vs CPU and WARP, hashes + walk unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Cold first-use per shader (queue ms): EvaluateFarField 1757, EmitNear 2429,
FarConnectorRings 784, FarSkirts 507+975; the far EmitCS dropped 2327 -> 3.2
by splitting out the near midpoint fallback into EmitNearCS. Warm-up under
0xFF00 ~2.0 s (worker count and thread priority do not change it; 0xFFFF run:
1.71 s). Whole view median CPU 1.428 s vs hybrid 3.154 s. Audits unchanged.
AMD DxCache was last written 2026-10-06 23:27 -- shaders may no longer be
cached across launches (unverified).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Top-down A/B/A controls: dispatching the 11 real compute shaders once removes
the 1.3-1.7 s first GPU Present (resource creation alone does not); 1 worker
also removes it and 8 workers bring it back. Default --gpu-far now prewarms
shaders asynchronously during shape compile (--gpu-far-no-prewarm = control).
The warm-up queue itself takes ~2.75 s, so whole view is unchanged:
CPU 1.354 s vs GPU 3.559 s (viewer clock), 1.473 vs 3.731 s (host clock).
Audits: 0 sign/topology diffs vs CPU/WARP, hashes + walk 5,642/5,642, survival.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Re-verified f2de8fe (hashes, walk 5,642/5,642, far topology 0 diffs vs CPU and
WARP, GPU survival). The remote LOD3/4 band was a direct-GPU coverage loss
(overlapping pixels identical, surface pieces missing, e.g. (6,2,0,lod3));
LOD3/4 now mesh on CPU, root cause in the GPU route still unidentified.
A minimal first-submit repro (pools, first-touch dispatch, device flags) did
not reproduce the 2.4 s stall. Whole view: CPU 1.438 s vs GPU 4.704 s median.
Crossing probe overflowed GPU output slots: no valid crossing latency.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna's session ended on an API "Bad Request" before a final report; state is
unverified by me. In the log: GPU clipped-skirt emitter (count/scan/emit,
protected signs and critical air honoured), exposed-skirt audit passing at
the end; first-present stall reproduced on every launch (2.36-2.59 s, not
exe identity/idle/CPU run) and hidden by holding the CPU-ready frame until
GPU residency. Remote LOD3-4 band still differs from CPU; draw ranges and the
one inspected LOD3 tile match, cause not yet found (one-tile probe mid-fix).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Resumed round: checkpoint rebuilt and re-verified (parity 3,436/3,436, walk
5,642/5,642, W1/W8 hashes 44/44, far topology 0 diffs vs CPU and WARP, far
survival PASS). Remote GPU holes were missing clipped LOD skirts: tiles with
visible skirt faces now fall back to the CPU mesh path, leaving only 16 far
tiles GPU-rendered until a GPU skirt emitter exists.
My interleaved runs (SkirtSeamVerify, process start -> whole view):
CPU 1.522/1.488/1.430 s; GPU 4.271/1.749/1.659 s. A ~2.8 s first Present
stall hit only my first GPU launch; Luna saw it on every fresh launch, so the
trigger is still unidentified. Headless CPU clipmap 0.789 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Luna measured the startup density-parity check at 1,049 ms, nearly all of the
~1.13 s before shape compile; it now runs after first display and still passes
3,436/3,436. CPU hashes 44/44 at 1 and 8 workers, walkability 5,642/5,642,
headless 8-worker clipmap 0.769 s. Stopped by the owner mid-round, while
auditing the remote far holes; first-launch stall and hybrid speed untouched.
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
Fr0zkaandClaude Opus 5.5