GameCult/Fensalir
Watch
9
0
Fork
You've already forked Fensalir
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
Fensalir/state/scratch.md

12 KiB
Raw Blame History

Scratch

Current Slice

Leap/Mimir point-cloud bridge: Mimir now declares a Leap disparity-derived FieldMesh PointList. Fensalir now preserves mesh SourceUri, allocates derived-from:* mesh buffers as UAV-capable GPU resources, fills those vertices/indices from the disparity SurfacePage in D3D12PointCloudFromDisparityCS, and renders the standard PositionNormalUvColor point list through D3D12PointCloudPS into the scene candidate targets. This is the first visible point-cloud lane; full SGM, metric calibration, splat sizing, and the global residual/calibration owner are still separate cuts.

Turn the native-domain reservoir from a native-resolution cache into a budgeted spatiotemporal sampler. The work grid now defaults to half present resolution and is clamped below native presentation scale; scene evidence, reservoir candidate/history, resolve, bloom, and match-window graph targets use that grid. The final present buffer is the only full-resolution surface.

The first comparison set is artifacts/fensalir-captures/reservoir-compare-20260529-231904-*: native versus texel-baseline changed 10.8208% of final-frame pixels, 0.0165% of rejection-debug pixels, and 0.2976% of shift-debug pixels. The repeatable measurement script is scripts/measure-reservoir-mode-comparison.ps1. First half-scale live capture: artifacts/fensalir-captures/budgeted-reservoir-20260529-234039-final.png at 1264x681 present output, nonblank by sampled pixel probe.

Temporal sequence harness smoke: scripts/capture-reservoir-mode-sequence.ps1 and scripts/measure-reservoir-sequence.ps1 produced artifacts/fensalir-captures/reservoir-sequence-20260529-234414-* at 640x360, final color only, ready frames 2 and 4. Native/baseline mode-delta changed 82.9025% of sampled pixels at f2 and 20.0206% at f4; temporal-native changed 78.5581%, temporal-baseline changed 48.9741%. These are probes, not a final ghosting score.

Masked sequence smoke: artifacts/fensalir-captures/reservoir-sequence-20260529-235031-* at 320x180, ready frames 2 and 3, final plus rejection debug. The sequence capture is now resumable: existing PNGs are skipped and still written into the manifest. Rejection-mask rows covered about 41% of frame pixels. Masked native/baseline mode-delta changed 87.4393% at f2 and 94.7517% at f3; masked temporal-native changed 83.2208%, masked temporal-baseline changed 83.4813%. This proves the mask path is live but still needs a high-budget reference or explicit disocclusion mask before it can claim ghosting quality.

Disocclusion owner signal: Debug mode 19, Reservoir Disocclusion, is derived from temporal rejection codes in D3D12Post.hlsl: previous UV outside history, previous history closer than expected travel, and support-overlap loss. The sequence measurer prefers disocclusion masks over broad rejection masks. Native-only smoke artifacts/fensalir-captures/reservoir-sequence-20260530-000151-* at 320x180, ready frames 2 and 3, final plus disocclusion debug, produced a mask covering 3.1098% of the frame and a masked temporal-native delta of 46.8867%.

Higher-work-grid reference: scripts/capture-reservoir-mode-sequence.ps1 accepts -ReferenceFieldReservoirScale, captures native-domain reference frames at that larger scale, and labels them reference in the manifest. scripts/measure-reservoir-sequence.ps1 reports reference-error-* rows against matching ready-frame counts, using disocclusion masks when present. Smoke artifacts/fensalir-captures/reservoir-sequence-20260530-001046-* compared native scale 0.5 against reference scale 0.75 at 320x180, ready frame 2, final plus disocclusion. Whole-frame reference-error-native changed 53.9240%; masked reference-error-native covered 3.0468% of frame pixels and changed 42.0368%.

Budget-pressure probe: scripts/measure-reservoir-budget-pressure.ps1 reads a sequence manifest with a same-time reference capture and ranks fixed tiles by masked candidate versus reference final-color error. It uses the candidate/reference disocclusion masks when available, falling back to rejection masks. Smoke artifacts/fensalir-captures/reservoir-budget-pressure-20260530-001046.* measured 1306 of 42864 pixels and found the top 5 of 50 tiles carried 99.3305% of measured reference-error delta. This is offline allocator evidence, not a runtime scheduler.

Runtime spatial reuse budget: GraphicsSettings.FieldReservoirSpatialReuseBudget now flows through renderer UI, runtime options, CLI/env overrides, headless capture scripts, frame constants, and D3D12ReservoirHistoryUpdateCS. Current and temporal reservoir rows are still written for every work-grid pixel; the budget only decides whether that pixel spends the 3x3 spatial neighbor sampling pass this frame. Default is 0.5. Smoke artifacts/fensalir-captures/reservoir-sequence-20260530-003006-* used native scale 0.5 / spatial budget 0.5 against reference scale 0.75 / spatial budget 1.0. Masked reference-error-native changed 63.7195% over 3.8261% of frame pixels, and reservoir-budget-pressure-20260530-003006.* found the top 5 of 50 tiles carried 90.8321% of measured reference-error delta.

Spatial budget debug: Debug mode 20, Reservoir Spatial Budget, visualizes the sampler branch that spends 3x3 spatial reuse. Green pixels spend spatial reuse; blue pixels skip it for the current frame. Smoke artifacts/fensalir-captures/reservoir-sequence-20260530-003746-* captured the mode at scale 0.5 / spatial budget 0.5. A bitmap probe counted 32078 green, 6818 blue, and 3968 other/filtering pixels in the 304x141 present output.

Window resize fix: The Fensalir window resize path now explicitly clears and flushes the D3D11-on-12 overlay context after disposing DirectWrite overlays and wrapped backbuffer resources, before IDXGISwapChain::ResizeBuffers. Resize also disposes program-output shared textures before recreating size-dependent output surfaces. Probe scripts/dev-reload.ps1 visible slot 20260530-215113-639e6612 survived seven programmatic resizes and logged rebuilt D3D12 targets at each size with no stderr.

Reservoir loss curve: scripts/measure-reservoir-loss-curve.ps1 now reports RGB reconstruction loss against same-time higher-work-grid native references. Metrics include MAE, MSE, RMSE, PSNR, changed pixels, max channel delta, and optional masked loss when disocclusion/rejection masks are present. Smoke artifacts/fensalir-captures/reservoir-sequence-20260530-221043-* compared native scale 0.5 / spatial budget 0.5 against reference scale 0.75 / spatial budget 1.0 at ready frames 1, 2, 4, and 8. Whole-frame PSNR was 21.668, 20.981, 20.911, and 21.690 dB respectively; RMSE was 21.043, 22.777, 22.962, and 20.992. This is a higher-budget online reference, not offline ground truth.

GPU pass timing: The D3D12 renderer now owns fixed pass-level GPU timestamp queries in per-frame resources and reports once-per-second averages beside CPU capacity logs. The timed passes are frame command recording, GPU sensor fusion, stereo depth, fractal reservoir update, tube-field update, height field, scene candidate, field-reservoir resolve, reservoir-history update, bloom, and presentation resolve. A 320x180 headless timing smoke after the reconstruction pass logged record 1.338 ms, scene-candidate 0.364 ms, reservoir-resolve 0.004 ms, reservoir-update 0.853 ms, bloom 0.029 ms, and presentation 0.077 ms over 61 frames in artifacts/dev-reload/fensalir-capture.out.log.

Reservoir presentation reconstruction: D3D12ReservoirPresentationResolvePS now performs the first final-pass spatiotemporal reservoir reconstruction filter. The reservoir/history grid remains below native resolution; the final present pass is the only full-resolution surface. Reconstruction samples a bounded 3x3 neighborhood from the resolved reservoir and weights neighbors by field id, travel/depth, normal, coverage, detail, color, guide confidence, domain validity, and neighbor agreement. The filter is maturity-gated so the first visible frame cannot blur from unearned history. Debug mode 21 visualizes reconstruction mix, confidence, and domain validity.

Latest measured curve: artifacts/fensalir-captures/reservoir-sequence-20260530-232852-* compared native scale 0.5 / spatial budget 0.5 against reference scale 0.75 / spatial budget 1.0 at 320x180, ready frames 1, 2, 4, and 8. Whole-frame RMSE/PSNR: f1 21.788 / 21.366 dB, f2 21.068 / 21.658 dB, f4 21.473 / 21.493 dB, f8 20.556 / 21.872 dB. Against the pre-reconstruction curve reservoir-sequence-20260530-221043-*, f2/f4/f8 improved and f1 regressed; average RMSE dropped from 21.944 to 21.221. This is competent enough to keep, but the next quality cut should target the first-frame regression and plot the curve against GPU pass timings in one artifact.

Hot Lesson

Area ReSTIR confirms that reuse over an area domain is invalid unless the reservoir stores the selected sample's actual domain coordinates and support. TSR confirms that temporal reconstruction must reject, reconstruct, and expose history trust at the layer that owns visible output. Fensalir needs both, but one level higher: selected samples are field and radiance claims lowered from semantic domains and a dynamic contribution tree into bounded GPU reservoirs.

The current nine-lane row ABI now carries first-class selected domain/sample-coordinate authority. A row keyed only by pixel, travel, field id, guide, proposal, and stats is no longer accepted as valid reservoir evidence. That authority now lives on a budgeted work grid, not on the final present grid: native-domain state describes selected field samples, while final resolution is reconstructed from bounded evidence instead of allocating native-res reservoirs. TubeField row samples store selected UV, logical column, curve coordinate, support footprint, domain kind, and shift kind; field id carries physical rolling-column identity. The old single-batch replay binding has been replaced by a bounded producer-keyed manifest plus raw source descriptor table, so the shader picks the replay source per selected TubeField sample instead of accepting/rejecting the whole frame by batch count. SDF history reconstructs the previous selected world hit from stored UV/travel, carries it through object center motion, and validates current ray/travel/support before temporal reuse can merge it. TubeField batches beyond the manifest source cap remain intentionally non-authoritative. Swapchain resize correctness depends on releasing both D3D12 backbuffer owners and D3D11-on-12 overlay references before ResizeBuffers. Disposing wrapped resources is not enough; the D3D11 immediate context must be cleared and flushed so it stops holding hidden references. Reservoir quality must be golfed against loss-over-time, not single-frame smoke deltas. The current curve is noisy and not monotonically improving, which means the candidate/reference comparison is exposing temporal instability rather than a settled convergence story.

Next Bounded Move

Follow docs/perfect-fensalir-machine-roadmap.md Phase 1:

  1. Put loss and GPU timing into one benchmark artifact, then tune the presentation reconstruction first-frame maturity gate so f1 no longer regresses while f2/f4/f8 keep their improvement. Do not add native-resolution reservoir intermediates.
  2. Add exact SDF producer replay only when client SDF distance functions have a shared replay include or manifest instead of duplicating shader policy in post.
  3. Add adaptive update/sample budget allocation only after the leakage harness can tell whether extra samples buy visible stability.
  4. Replace the bounded TubeField replay source descriptor table with a replay atlas only if real scenes hit the cap.
  5. For Mimir geometry, verify live Leap point clouds in the editor, then add the residual/calibration owner that consumes Leap point error before other cameras contribute motion/RGB evidence.

Verification

  • dotnet build Fensalir.sln
  • dotnet test tests/Aquarium.Engine.Tests/Aquarium.Engine.Tests.csproj
  • dotnet test tests/Aquarium.Engine.Fractal.Tests/Aquarium.Engine.Fractal.Tests.csproj
  • .\scripts\dev-reload.ps1 -Headless -RetainSlots 4 -ClientProject src\Aquarium.Fensalir\Aquarium.Fensalir.csproj