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
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
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
Whole far field on the RX 7900 XTX: 112 ms total (connector rings 61 ms). Cold hybrid 2.303 s vs
all-CPU 2.314 s (within noise): the far field is no longer on the wall-clock critical path.
0 sign/topology differences vs CPU and WARP incl. 127,401 protected and 353 thin-feature cells;
repeat GPU runs identical; far survival passes on GPU meshes; CPU hashes unchanged.
Open: CPU/GPU vertex distance mean 1.15 mm but max 64.56 cm, 247 tiles over the quantisation bound.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Up to 32 tiles per deferred command list, one polled event per batch, no per-tile waits; CPU
upload/enqueue 723 -> 46 ms. GPU far meshes pass far survival (7,200 strata boundaries, 10,976
connectors, 0 missing/holes). But hybrid cold 4.36 s vs CPU 2.29 s: whole-tile GPU timestamps sum to
646 ms (0.58 ms/tile; step 3's 0.03 ms covered part of the work only), divergent per-thread shape
loops; CPU far-input export 1.6 worker-s. CPU 44/44 hashes unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
1,114 of 1,266 far tiles meshed on the GPU; vs CPU and vs WARP: 0 sign/topology differences,
including the protected/thin exception audit; max symmetric vertex distance 0.125 cm; repeat RX runs
identical; GPU field+mesh 0.029 ms/tile average. Strata survival 7,200/7,200, but connector survival
fails (GPU lacks the connector repair panels). Hybrid cold clipmap 5.74 s vs CPU 2.23 s: dispatch and
readback overhead, not compute (~33 ms of GPU work in total).
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
Fr0zkaandClaude Opus 5.5