Skip to content

FLUID_BOUNDARIES.md

Latest commit

 

History

History
123 lines (107 loc) · 7.65 KB

FLUID_BOUNDARIES.md

File metadata and controls

123 lines (107 loc) · 7.65 KB
Raw

Substep boundaries and error-triggered density repair

This continues the adaptive-fluid implementation at 7f89656; it is not a completed multiresolution MAC solver or narrow-band fluid system.

Integration

  • FluidColliderTimeline owns only a bounded array of rigid colliders. It interpolates translation and shortest-arc rotation from the last simulated pose to the current render pose, with an immutable upload slice per substep. FluidSystem selects each slice through the existing root SRV and rebakes the solid MAC boundary. No particle readback, new queue, or frame wait is introduced.
  • Boundary velocities use elapsed simulation time. Zero-substep frames retain the simulation anchor; paused placements, resets, and geometry replacement establish new anchors. Small-angle angular velocity uses atan2, avoiding acos(w) quantization. The final SDF transform exactly matches the renderer.
  • The existing reconstruction, importance, resampling and ownership systems use the same FluidGpuView collider address. After simulation it points at the exact current pose used by the canonical scalar field and procedural DXR.
  • Particle contacts are resolved before positional density projection as well as afterwards. The density solve sees displaced water rather than water still inside a newly moved solid. The full unilateral density error is used, with RHS bounded at 0.5 and the existing independent 0.2h displacement trust region.
  • One additional density correction is requested entirely on GPU if a current occupied nonsolid cell has kernel density (including missing solid support) above 1.02. Generated indirect arguments control Jacobi, particle displacement, contact and rebinning. Empty requests skip their shader work. Classification and the error gather are still uniform, and a triggered solve remains globally coupled; this is not yet a spatially compact or converged nonlinear solver. Corrected bins are retained for the next substep's P2G, avoiding a redundant histogram/scan/scatter when no particle positions changed between those passes.

The need to incorporate obstacle displacement into density correction is also visible in the author's IDP-FLIP reference implementation. Our bounded contact/projection splitting is original engine-integrated code, not that reference's full boundary formulation. Cut-cell solid capacities, consistent swept-volume fluxes, two-way rigid coupling and CFL/CCD handling of arbitrary fast objects remain separate work. Interpolation alone is not CCD.

Diagnostics

maxRelativeDensity retains the original pre-first-correction measurement. Validation-only DensityMeasure computes postStepMaxRelativeDensity from the final particles and coarse owners after contacts and ownership exchanges. It does not clear pressure or alter simulation state. Reports also include peak cell coordinates/solid distances, post-step nonsolid peak, occupied volume, and whether the final substep requested a repair.

Both peak locations are checked by an independent double-precision CPU gather over current samples and the separable owner measure, not over GPU bins. meanHeight now includes owner mass and position; particleMeanHeight explicitly reports the particle-only centroid. These expensive checks run only during bounded validation, not normal play or profiling.

Investigation record

The 180-frame sphere fixture squeezes water against the right domain wall. It exposed real nonsolid compression, not a checkerboard shading artifact or merely stale reporting. Matched uniform-particle and owner runs, RTX 5090, 100k rest mass:

Intermediate implementation Uniform post-step peak Owner post-step peak
Original stepping + new audit 2.60107 2.81103
Substep-aligned colliders only 3.04208 3.13507
Contacts before density too 2.47046 2.53471
Full bounded density RHS too 1.06713 1.11644
GPU error-triggered repair too 1.01246 1.00964

These are intermediate correctness experiments, not benchmark speedups. In particular, collider interpolation alone did not cure compression. The independent peak audit discrepancy in the last row was below 1e-6. The original pre-correction density gate is retained, with an additional post-step comparison; no reference threshold is relaxed to hide a regression.

Validated checkpoint

  • All 15 interior/reference cases pass, including the formerly failing moving sphere, 600-frame calm run, forced-fine/pause/single-step, room emitter, falling water, multigrid, passive bulk, FG and ReSTIR PT integration. Calm water retains 96 owners / 6,624 mass units and requests no extra density solve; the moving sphere restores all 96 owners and requests repair. Issued mass remains 100,000.
  • The final sphere run ends at pre-correction density 1.01375 and post-step 1.00962. The matched uniform reference ends at 1.01056 and 1.00713. Both unchanged pre-correction and new post-step relative gates pass, as does a new absolute 5% post-step compression limit. The resampled-only reference still differs from uniform water and is not substituted for it to obtain a pass.
  • 75 Node tests and seven Windows CTests pass. CPU timeline tests cover skipped render frames, pause/reset, rigid midpoint/final endpoints, noncommuting rotations, slow rotation and quaternion hemisphere crossing.
  • Six ordinary GPU regression cases pass: constant, affine, compression, roundtrip, controls and the 2,400-substep APIC run. Additional moving-SDF render and surface-control cases pass, including canonical DXR roots and photon checks.
  • The 320-frame room orbit/rolling sequence passes --preserve against the prior water-temporal-interior sequence: maximum brightness change 0.321%, raw motion-delta increase 0.653%, RR motion-delta increase 2.566%, one RR reset. This is not pure Monte Carlo variance or proof of general dynamic ownership optical convergence; the shallow room has no eligible owners.
  • Normal Windows Release was tested. GPU-based validation was not run: Windows Graphics Tools was previously unavailable (0x887A002D).

Three serial final room/orbit runs, wall inlet and whitewater enabled, 1080p Balanced, FG off, no invariant snapshots: 804 measured frames have median raw GPU frame time 12.567 ms, p95 15.567 ms, simulation 2.957 ms, and reconstruction 1.159 ms. This is the current operating point, not a controlled speedup comparison against the old compression-buggy solver. The correctness repair has a cost; fully sparse/local projection remains the optimization path.

boundaries-validation.json records fresh bounded reports; the independent image comparison is preserved in boundaries-temporal.json. Reproduce with test-fluid-interior.ps1, the scoped test-fluid.ps1 / test-fluid-render.ps1 cases, test-water-temporal.ps1 -Name water-temporal-boundaries -Whitewater -Adaptive -Interior, and validate-water-temporal.mjs <runtime> water-temporal-boundaries water-temporal-interior --preserve. record-fluid-boundaries.mjs reads those reports without launching tests and rejects reports older than the executable. Current timing runs use profile-room.ps1 -Tag boundaries-final -Repeats 3 -CaseFilter '^orbit$' with profiling run serially and without simultaneous builds or invariant tests. The old interior-validation.json remains a historical record of the failed 7f89656 checkpoint.

General cut-cell volume conservation, flowing coarse ownership, multiresolution MAC coupling and budget-controlled spatial work remain outstanding. A bounded extra projection is not a guarantee that arbitrary impacts converge below 2%.