44r0n7/deepgrain-foundry
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Multi-core simulation: grain movement and heat across a worker pool (t88 phase 3)

Browse Source 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
44r0n7
2026-10-28 03:23:43 -04:00
co-authored by Claude Opus 5.5
parent 20963c776e
commit 7468d00efb
35 changed files with 578 additions and 113 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/agent-log.md
+82
View File
@@ -4,6 +4,88 @@ What agents have done, newest first. Add an entry for every piece of work: the d
note or request it was for, what changed, the files touched, the tests run, and anything left
open or not verified. Keep each entry short; the git history has the details.
## 2026-10-28: t88 phase 3 resume, final review and verification (Codex)
- Preserved the interrupted uncommitted implementation and reviewed the schedule, shared
writes, deferred effects, startup gate, test changes and profiler against the Phase 3 brief.
- Reproduced a deferred-liquid bug before fixing it: replaying liquid() rolled the movement
chance twice. Extracted liquidSearch() so the deferred search resumes at its exact direction
and distance without another random draw; removed the now-unneeded RNG replay records.
Added a failing-then-passing regression and a long water shelf to the repeated 320-tick
inline/1/2/8-worker and reversed-order fixture.
- Chunk passes now skip per-cell renderDirty stores in wakeXY/wakeHeatXY; the scanned chunk
and write halo provide their invalidation. Outside passes, explicit invalidation remains.
- All headless suites passed. Initial sandboxed browser gate failed (no GPU device); an
elevated capability check confirms RTX 5070 WebGPU, isolation and 24 cores. Firefox still
has no adapter and correctly requires WebGPU. Final host-GPU browser gate in progress.
- Files additionally changed on resume: sim/{cells,workers,world}, test/t_threads.js,
PROJECT_MAP and tracking. Full fresh profile, two-minute live run and screenshot inspection
pending. No commit; DESIGN and owner's Context untouched.
## 2026-10-27: t88 phase 3, multi-core simulation (Codex; verification in progress)
- SharedArrayBuffer allocation for world arrays (ArrayBuffer only when unavailable in Node).
Four (chunkX&1, chunkY&1) movement passes then four heat passes. Each chunk scopes RS to
hash(world seed,tick,chunk,pass), restoring the global machine/worldgen RNG afterwards.
Long liquid reads outside a 24-cell horizontal halo are deferred and replayed in chunk order;
movement writes reach at most 8 cells, so parallel read/write regions cannot overlap.
- Pass-local wake/heat flags merge by maximum; per-chunk heat outputs merge first. Discoveries,
events, fuel ignition counts, wheel flow, cloud moisture, broken machines and partLost merge
in chunk order. Flyers/launchGrain remain reachable only from main-thread transport, not these
cell passes; their existing state is included in the determinism hash. Direct shared writes
are disjoint cell/powder outputs or idempotent render/roof bits.
- Browser Blob pool uses min(16,hardwareConcurrency-2,setting), at least one. Main thread owns
machines, transport and game logic, processes a batch in each pass, then polls Atomics.load.
No Atomics.wait on the document: it is prohibited by
https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Atomics/wait
and https://tc39.es/ecma262/2024/multipage/structured-data.html. Polling is synchronous and
occupies the main core; it preserves atomic ticks so input/save/render cannot race workers.
No whole-simulation worker or asynchronous rendering overlap. Worker timeouts/errors stop
frame simulation. Small jobs run inline and larger batches use only enough helpers to
amortize dispatch (32 movement or 8 heat chunks per helper). Pool setting applies on load.
- Invalidation: remove swap's duplicate hot marks, gas marks and repeated coordinate-to-chunk
function calls. Each scanned movement chunk marks its halo once; heat marks conduction edges.
wake functions reuse their already computed chunk index for one direct render bit store;
rare swaps/building/fog/dig/background writes retain explicit marks. Paused clean draws upload
zero world cells. GPU equality passes without changing any tolerance.
- Measurements so far: fresh Phase 2 settled/factory/heat 3.537/4.077/5.824 ms/tick vs frozen
baseline 3.715/3.657/4.938. Isolated removal of hot dirty calls in a temporary Phase 2 copy
gave 3.530/3.921/5.936: it did NOT recover the full 15% on its own. Do not claim it did.
Initial unbatched pool factory 4.21; batched trial 3.04 ms/tick. Full final matrix pending.
- Final full matrix (`/tmp/t88-phase3-final.json`, `/tmp/t88-phase3-final-perf.txt`, 218.0 s),
RTX 5070, 24 logical cores, 1920x1080, 2048x1024 world. Pool 16 helpers, peak active 4
(settled) / 5 (factory, heat), plus main-thread batch. All values below are ms.
settled: baseline 3.715 -> wall/main occupied 1.897 (1.96x); main compute excluding polling 1.624, polling 0.272.
factory: baseline 3.657 -> wall/main occupied 2.495 (1.47x); main compute excluding polling 2.037, polling 0.458.
heat: baseline 4.938 -> wall/main occupied 2.649 (1.86x); main compute excluding polling 2.236, polling 0.412.
Real-loop CPU work p50/p95, zoom 2 / zoom 6; pixel then smooth:
settled: pixel 7.08/8.58 / 2.77/3.83; smooth 8.46/9.98 / 2.98/4.13.
factory: pixel 8.48/10.05 / 4.96/5.95; smooth 9.59/10.80 / 5.13/6.12.
heat: pixel 8.53/10.56 / 4.88/6.03; smooth 10.05/11.51 / 5.17/6.21.
Frame intervals p50/p95 are 16.67/16.67 for all these runs (60 fps). Tick wall time and
occupied main-thread time are equal: polling blocks this core; rendering does not overlap.
The final threaded sim meets the baseline target in every scene; standalone dirty-only
recovery remains unproven/unmet (numbers above). Recommend a future whole-simulation worker
to remove the remaining 1.90-2.65 ms from the main thread; it would require an explicit
ownership/input/save boundary and is outside this phase.
- Re-baseline: three worldgen material hashes and seeded pile bounds (57x17 -> 46x19 grains).
The heat oracle retains its unoptimized per-cell computation, updated for buffered pass
outputs, and is compared against all four new passes. Same throughput assertions retained:
conveyor feed heap widened, downhill-screen warm-up lengthened; load-wake deadline kept at
eight ticks, fuel consumption checked later instead of demanding a random burn on that tick.
Frozen renderer PNGs and tolerances unchanged: original Phase 2 world bytes captured from the
Phase 2 simulation (only invalidation edits) in the temporary copy, compressed as golden-world fixtures, so changed
terrain doesn't get counted as changed renderer pixels.
- Headless suites pass; repeated inline/1/2/8 real-worker hashes pass (world arrays + JS effects).
Extended test passes 320 busy ticks including stamp wrap, reversed chunk order, Node fallback.
First full browser gate: gameplay 361, smooth 28, console 18, capabilities 7, GPU equality 12,
light 71 and Firefox 2 pass; golden sky failed due changed worldgen terrain. Frozen-input
golden rerun, final browser gate, full profile and 2-minute play-through pending.
- Files: sim/{world,cells,heat,workers,fog,machines,weather}, ui/{threads,title,state,settings,loop},
render/gpu/device, build.py/manifest; harness, threads/heat/load/corner/conveyor/screen/capability/
golden tests and fixtures; profiler, README, TESTING, PROJECT_MAP and t88 tracking.
No commit; DESIGN and owner's task context untouched. Phase 3 remains in progress.
## 2026-10-27: t88 phase 2d, required WebGPU and dirty uploads (Codex)
- Per-32x32-chunk `renderDirty` marks cell writes, heat with its conduction halo, stationary