flatten: heal degenerate cells so SLIM stops NaNing / crashing#988
flatten: heal degenerate cells so SLIM stops NaNing / crashing#988SuperOptimizer wants to merge 1 commit into
Conversation
…IM stops NaNing
Full-res (and even lightly-decimated) SLIM flattening of some segments
failed: -nan at iter 1, or a PaStiX heap corruption / SIGSEGV. Root
cause was a chain of three issues, all stemming from degenerate input
geometry the tracer emits:
1. Collapsed grid cells. The grower can leave grid-ADJACENT points at
the same (or sub-voxel) 3D location (~6k scattered cells in the repro
segment). Healthy adjacent cells are ~1/scale voxels apart. These make
zero-area triangles whose symmetric-Dirichlet energy is infinite ->
singular system -> NaN at iter 1.
Fix: healDegenerateCells() (core/InpaintSurface) drops connected
clusters of collapsed cells (dilated by a ring) and refills them via
inpaintSurfaceHoles, iterating a few passes. Threshold is relative to
1/scale so it adapts across segments. Optional dropped_mask lets
callers clear the same cells in other channels (approval, etc.).
2. Unrepairable stragglers. A few fully-collapsed 2x2 blocks can't be
separated locally, and even one zero-area triangle re-NaNs SLIM.
Fix: a scale-relative emit guard in vc_tifxyz2obj skips any triangle
with a collapsed edge or sliver altitude -> guarantees 0 zero-area
triangles reach the OBJ (worst case: a sub-cell hole).
3. Orphan vertices. Skipping triangles can leave a vertex referenced by
no surviving face. An orphan is an all-zero row/col in L = AtWA, which
crashes the PaStiX Cholesky (double free / SIGSEGV in factorization,
surfacing as an Eigen SparseMatrix free). This is what looked like a
solver bug.
Fix: vc_tifxyz2obj now emits in two passes -- collect surviving
triangles first, then emit only the vertices they use. No orphans.
Also set spm.replicated = 1 in the Pastix6LLT wrapper: spmInit leaves it
-1 ("uninitialized"), which steers PaStiX down distributed-matrix code
paths for our single local matrix. Latent correctness fix.
Verified on 20230702185753_v11 (3.9M verts, ~6k collapsed cells): was
-nan at iter 1 at every decimation; now flattens cleanly via the VC3D
GUI and via CLI at keep=2/6/25 (0 orphans, converged L2 ~1.0, valid OBJ).
|
Someone is attempting to deploy a commit to the scroll Team on Vercel. A member of the Team first needs to authorize it. |
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b7285d5fb5
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const double scale = (std::isfinite(sc[0]) && sc[0] > 0.f) ? sc[0] : 0.0; | ||
| const int healed = vc::core::util::healDegenerateCells(points, scale); |
There was a problem hiding this comment.
For surfaces whose metadata has non-square scale (scale[0] != scale[1]), this passes only the x scale into healDegenerateCells, whose 4-neighbor checks use that single expected spacing for both horizontal and vertical edges. When the y spacing is legitimately smaller than 1/scale[0], healthy vertical neighbors can be classified as collapsed and dropped/inpainted before export, potentially corrupting large anisotropic segments; the emit-time expected_step below repeats the same x-only threshold. Please carry both scale components through the degeneracy checks or use direction-specific thresholds.
Useful? React with 👍 / 👎.
2 participants
Problem
Full-res SLIM flattening of some segments failed hard:
-nanenergy at iter 1 (and, once that was worked around, a PaStiX heap corruption / SIGSEGV during factorization). It happened at every decimation level, so decimation never rescued it.Root cause is degenerate input geometry the tracer emits: grid-adjacent points collapsed onto the same (or sub-voxel) 3D location. A healthy sheet keeps adjacent cells ~
1/scalevoxels apart; the repro segment (20230702185753_v11, 3.9M verts) had ~6k coincident adjacent points scattered in ~2,500 tiny clusters. These make zero-area triangles → infinite symmetric-Dirichlet energy → singular system → NaN.Fix (three layers, all geometry-side)
healDegenerateCells()(core, next toinpaintSurfaceHoles): detects collapsed grid cells, drops the connected cluster (dilated by a ring) so the inpaint gets a clean interior hole with healthy boundary, and refills viainpaintSurfaceHoles, iterating a few passes. Threshold is relative to1/scaleso it adapts across segments. Optionaldropped_maskout-param lets callers clear the same cells in other channels (approval, etc.). Repairs ~70k cells on the repro.Scale-relative emit guard in
vc_tifxyz2obj: a few fully-collapsed 2×2 blocks can't be separated locally, and even one zero-area triangle re-NaNs SLIM. The guard skips any triangle with a collapsed edge or sliver altitude → 0 zero-area triangles reach the OBJ (worst case: a sub-cell hole the flattener tolerates).Two-pass emit in
vc_tifxyz2obj: skipping triangles can leave a vertex referenced by no surviving face. Such an orphan is an all-zero row/col inL = AᵀWA, which crashes the PaStiX Cholesky (the "double free / SIGSEGV" that looked like a solver bug). Now we collect surviving triangles first, then emit only the vertices they use — no orphans.Plus a one-line latent fix in the
Pastix6LLTwrapper:spmInitleavesspm.replicated = -1("uninitialized"); set it to1(single local matrix) so PaStiX doesn't take distributed-matrix branches.Verification
20230702185753_v11: was-nanat iter 1 at every decimation. Now flattens cleanly via the VC3D SLIM panel (GUI, decimated path, new segment created) and via CLI atkeep=2/6/25— 0 orphans, 0 zero-area triangles, converged L2 ≈ 1.0, valid output OBJ.Notes