Repository navigation
The SAT's edge table holds every edge a hull can have - #101
Merged
Merged
Conversation
The edge-pair phase gathers A's edges into a per-worker table of 14 field
arrays, sized for 64 edges ('MAX_HULL_EDGES / 2'). MAX_HULL_EDGES counts
full edges -- the hull builder allows 2 * MAX_HULL_EDGES half-edges, 128
edges -- so a hull of more than 64 edges (a 32-sided cylinder has 96, a
hull of 40 points on a sphere about 114) wrote its later edges over the
earlier ones' fields and, past about 61, beyond the worker's block into
the next worker's or past the allocation. The edge axes came out wrong:
missed or ghost edge contacts, wrong separations in the SAT cache. And
#94 kept B's normals and vertices in the first clip polygon's scratch on
the same mistaken bound (66 of them; there can be 256).
The table holds MAX_HULL_EDGES edges now, its field offsets derived from
that; B's normals and vertices go in room after it in the same block, so
the allocation count is unchanged.
test_manifold's test_many_edges: two 32-sided cylinders (96 edges each) at
300 random poses, the SAT's best edge separation against brute force over
every edge pair -- 0 of 300 agreed before, all but at most 3 now. Found by
an audit of the manifolds against Box3D e77352c, otherwise faithful.
…s constants A constant built from another module's constants (hull.MAX_HULL_FACES + hull.MAX_HULL_VERTICES) is expanded where this module is imported, and test_mesh and test_height_field import the manifold without the hull: 'Undefined variable hull'. The sum is written out (256) with the reason.
This was referenced Oct 5, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
I audited contact manifold generation against Box3D
e77352c: the convex pairs, triangle manifolds and mesh contacts. All of it matches upstream except this one defect, which corrupts memory.The bug: the SAT's edge-pair phase gathers A's edges into a per-worker table of 14 field arrays. The table was sized for 64 edges, on the assumption that
MAX_HULL_EDGEScounts half-edges. It counts full edges: the hull builder allows up to 128 edges (256 half-edges).A hull with more than 64 edges (a 32-sided cylinder has 96, a 40-point sphere hull about 114) wrote its later edges over the earlier ones' fields. Past about 61 edges it also wrote beyond the worker's block, into the next worker's block or past the end of the allocation. The resulting edge axes were wrong: missed or ghost edge contacts, and wrong separations stored in the SAT cache.
#94 made the same mistake: it put B's normals and vertices in the first clip polygon's scratch, assuming at most 66 of them. There can be 256.
The fix: the table holds
MAX_HULL_EDGESedges, with its field offsets derived from that size. B's normals and vertices get room after the table in the same block, so the number of allocations is unchanged.test_manifold's newtest_many_edgesplaces two 96-edge cylinders at 300 random poses and compares the SAT's best edge separation with a brute force over every edge pair. Before the fix, 0 of 300 agreed; now all but at most 3 do.All 28 test programs pass locally on aether 0.772.0, with no warnings. The machine was too loaded for meaningful timings.
bench/manifold's contacts are identical to main (same points and separation sum).