Stop generating the triangles the simplifier only has to take away - #124
Conversation
Dual contouring emits one vertex per surface cell, so a flat wall 64 voxels across becomes thousands of triangles that QEM then collapses back to two. Profiling for #111 put simplification at ~80% of an export at res 256, and 93% of the triangles it removed carried *zero* error — the mesher undoing its own over-generation. Octree vertex clustering decides, before any triangle exists, which cells can share a vertex: a node collapses when one point still fits every hermite sample underneath it within the budget. The reason this is safe is the point of the change. Adaptive dual contouring (Ju et al. 2002) contours the octree directly, which needs cellProc/faceProc/edgeProc recursion to stitch differently-sized neighbours and produces holes in an exported STL when that is wrong — which is why #111 was filed as "not small". None of that machinery is needed. Step 2 still emits exactly one quad per sign-changing grid edge over the same four cells; only the vertex those cells name changes. The output is a quotient of a mesh that was already watertight, so there are no cracks to patch. What clustering can break is manifoldness, when two surface sheets land in one node. A cell with more than one patch never joins a cluster, and — deliberately more conservative — it stays keyed in place so it fails every node above it too, keeping merged vertices out of regions holding geometry they do not describe. At res 256, the resolution exports actually run at: Wall Bracket 14529 -> 6550 ms 16478 -> 3274 tris maxdev 0.071 -> 0.071 Open-Top Enclosure 22884 -> 14438 ms 50694 -> 21164 tris maxdev 0.046 -> 0.046 Knob 15626 -> 9474 ms 68654 -> 58312 tris maxdev 0.016 -> 0.018 M3 Standoff 70504 -> 6546 ms 47710 -> 32256 tris maxdev 0.077 -> 0.077 Vent Grid 80850 -> 14614 ms 104548 -> 70708 tris maxdev 0.044 -> 0.045 Max deviation — what matters for a print — is unchanged on three and moves by 0.001-0.002 of a voxel on the two curved ones. Every output is watertight with zero non-manifold edges. The timings were taken on a machine with other work on it, so treat the two largest as unconfirmed; CI's bench job measures this cleanly. The triangle counts are deterministic. Off by default, and the dense path is verified bit-identical to before across every preset, with and without the active-block mask, so the geometry suite still asserts against exactly the mesh it always did. `cluster.test.ts` covers the collapse rule and then re-applies dualContour's own gates to the clustered mesh — watertight, edge-manifold, outward-wound, correct enclosed volume, vertices on the isosurface, and the two-sheets-in-one-cell case. A 10mm cube at res 32 comes out as exactly 12 triangles enclosing exactly 1000.00 units of volume, where the dense mesh spends 4800. The budgets move to `budgets.ts` because the benchmark had already drifted from the worker it claims to measure. Clustering gets half the allowance and QEM the other half: both stages move vertices and their errors add. Closes part of #111. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
🚀 Preview deployed: https://perf-octree-vertex-clusterin.sinter.pages.dev (updates on every push to this PR) |
Clean numbers from CI, replacing the caveated local onesThe local timings in the description were taken on a contended machine. CI's bench job ran this branch and Totals
The simplification stage alone, which is what this targets:
Correcting the description: the 10.77x on M3 Standoff and 5.53x on Vent Grid were load artifacts, exactly as flagged. The real range on this corpus is 1.09–2.20x overall. My res-256 measurements of 1.58–2.22x on the better-behaved presets line up with CI; the two outliers do not, and should be ignored. Contouring gets slower, which is the honest trade: Clustering costs work during contouring and saves more than it costs in simplification. Net positive on every model in the corpus, but the margin on Why the corpus gains less than the presets: the bench runs below export resolution, so simplification's share of the total is smaller than the ~80% measured at res 256. |
1 participant
Implements the approach prototyped in #111.
The problem
Dual contouring emits one vertex per surface cell, so a flat wall 64 voxels across becomes thousands of triangles that QEM then collapses back to two. Profiling put simplification at ~80% of an export at res 256, and 93% of the triangles it removed carried zero error — the mesher undoing its own over-generation.
Why this is not the risky version
#111 was filed as "not small" because adaptive dual contouring (Ju et al. 2002) contours the octree directly, needing
cellProc/faceProc/edgeProcrecursion to stitch differently-sized neighbours — and producing holes in an exported STL when that is wrong.None of that machinery is needed. Quads are already emitted per sign-changing grid edge, each referencing four cells through
vertexFor. If cells in a collapsed node hand back the same vertex, connectivity is untouched: still one quad per edge, still the same four cells. The result is a quotient of a mesh that was already watertight, so there are no cracks to patch.The real risk moves from cracks to manifoldness, when two sheets land in one node. A multi-patch cell never joins a cluster, and stays keyed in place so it fails every node above it too — deliberately more conservative than needed, to keep merged vertices out of regions holding geometry they do not describe.
Results at res 256 (what exports actually run)
Max deviation — the number that matters for a print — is unchanged on three and moves by 0.001–0.002 of a voxel on the two curved ones. Every output is watertight with zero non-manifold edges.
Caveat on the timings: these were taken on a machine running other work, and the two largest speedups (M3 Standoff, Vent Grid) have implausibly slow "before" numbers. Treat them as unconfirmed — CI's bench job measures this cleanly, and the interleaved best-of-3 at res 128 gave a steadier 1.5–2.2x. The triangle counts and deviations are deterministic and repeated exactly across runs.
The win is shape-dependent, which is expected: flat-faced parts collapse a lot, curved ones little. On the bench corpus,
enclosuredrops 14,420 → 7,494 triangles whilepatternedbarely moves.Safety
mainacross every preset, with and without the active-block mask — so the geometry suite still asserts against exactly the mesh it always did.cluster.test.tspins the collapse rule, then re-appliesdualContour's own gates to the clustered mesh: watertight, edge-manifold, outward-wound, correct enclosed volume, vertices on the isosurface, and the two-sheets-in-one-cell case.Two things found along the way
The benchmark had drifted from the worker it claims to measure.
bench/mesh.bench.tsreplicates the export pipeline rather than calling it, and was still timing a stage configuration the exporter no longer used. Budgets now live inbudgets.tsand both import them. A benchmark that measures something nobody runs is worse than none, because it is believed.Error budget is now split. Clustering gets half the allowance and QEM the other half, because both stages move vertices and their errors add — spending the whole budget in the mesher would let QEM spend it again on top.
Also filed
#123 —
fitPrimitivecosts ~5–9 s per fit, independent of grid resolution. That is a user-facing import cost, and it is also whyfitPrimitive.test.tsis repeatedly the first thing to time out under load. Not addressed here.🤖 Generated with Claude Code