[mfagerlund]

Two small changes to GeodesicAlgorithmExact (src/surface/exact_geodesics.cpp), one commit each.

1. max_propagation_distance is ignored when there are no stop points

The documentation says: "Once propagate identifies all vertices within max_propagation_distance of the source points, it stops." Without stop points it never stops early: check_stop_conditions returns false whenever m_stop_vertices is empty, so the bound has no effect and the propagation covers the whole mesh.

Reproduction:

GeodesicAlgorithmExact mmp(*mesh, *geometry);
mmp.propagate(mesh->vertex(0), bound); // a bound smaller than the mesh's extent
VertexData<double> dist = mmp.getDistanceFunction();
// on master, every vertex has a finite distance

Fix: with no stop vertices, check_stop_conditions returns true once the queue has passed a finite bound. With no bound and no stop points it still runs to the end. Vertices beyond the bound may now be left at GEODESIC_INF. Vertices closer than the bound keep their exact distance.

Behaviour with stop points is unchanged. When both a finite bound and stop points are given, propagation continues until both conditions hold. The documentation can be read either way for that case, so I left it alone.

2. Vertex Gaussian curvatures are required per popped interval

The saddle test in propagate called requireVertexGaussianCurvatures() and unrequireVertexGaussianCurvatures() for each endpoint of each popped interval. This change requires them once before the loop and releases them after it.

This is a cleanup rather than a large speed-up: unrequire() does not free the buffer, so the repeated calls cost a counter update each, not a recomputation.

Results

Measured with master's and this branch's exact_geodesics.cpp, on 14 common test meshes (1.8k to 112k vertices) and 10 Thingi10K models, 20 vertex pairs each.

Stop-vertex queries (propagate(v0, GEODESIC_INF, {v1})): 480 of 480 distances are bit-identical to master. There is no speed change to claim: the removed calls cost about 13 cycles per require/unrequire pair, against 2900 to 9600 cycles per popped interval, so about 0.3 to 0.9% of a query by that count. A master-vs-branch comparison could not resolve this on a loaded machine (geometric mean of CPU cycles master/branch 0.999; the same comparison between two identical builds gave 1.015, with per-mesh ratios of 0.87 to 1.25 and 0.93 to 1.15 respectively).

Bounded propagation without stop points (propagate(v0, d), with d the distance from v0 to v1, 3 pairs per mesh):

• Vertices reached: 0.53x master (geometric mean; 0.31x to 0.78x per mesh). Master always reaches every vertex.
• CPU cycles: master/branch 1.76 (geometric mean; 1.04 to 3.7 per mesh).
• Of the 853007 vertices closer than the bound, all match a full propagation bit for bit except one, which differs by 1.8e-15. That vertex is v1 itself, on the bound, and the difference is rounding.

On a larger model, Thingi10K 719791 (871k vertices), three nearby pairs: bounded propagation reached 1466, 4933 and 6985 vertices in 11 to 24 Gcycles, about what the matching stop-vertex query costs. Those three stop-vertex distances are bit-identical to master. Master's bounded run was not measured there, since it covers the whole mesh: a single stop-vertex query between distant vertices had not finished after 84 CPU minutes.

Tests

New test/src/exact_geodesics_test.cpp:

• MaxPropagationDistanceWithoutStopPoints: fails on master, passes here.
• MaxPropagationDistanceWithStopPoints: stop points alone, and stop points with a bound.
• ResultIndependentOfRequiredQuantities: the distance field is the same whether or not the caller holds the curvatures.

The full test suite passes (204 tests).