computeFaultDistances faked a nearest-point query with cvf::BoundingBoxTree,
which only supports box overlap: it grew a cell-sized box until it overlapped a
fault face. For cells far from any fault this did many full tree traversals and
then brute-forced the distance to a large, over-collected candidate set, and the
answer was only approximate. This was slow on grids with ~10 million cells.
The metric is the distance from each cell center to the nearest fault face
center, which is exactly a nearest-point search. Replace the bounding-box tree
and the grow-the-box loop with a nanoflann KD-tree built over the fault face
centers and a single nearest-neighbor query per active cell. A zero-copy adaptor
reads the existing std::vector<cvf::Vec3d> directly. The OpenMP loop is retained
since nanoflann queries are read-only. The result is O(F log F) build plus
O(N log F) queries, the full grid-node array copy is gone, and distances are now
exact rather than approximate.
Add nanoflann as a vcpkg dependency, wired in like Clipper2 as an imported
target.
Refs: https://github.com/OPM/ResInsight/issues/14128
The existing FAULTDIST result always considered every fault in the main
grid. Users with many faults need to compute distance fields against a
named subset, so this adds a Fault Distance Results collection under
each view's Faults node. Each entry holds a multiselect of faults and a
name (FAULTDIST1, FAULTDIST2, ...) and publishes the result into the
Generated cell-result category.
The per-cell BVH-based distance loop was extracted from
RigFaultDistanceResultCalculator into a reusable utility that accepts
the subset of faults to include; the original all-faults entry point
delegates to the same utility and keeps the static-native FAULTDIST
behavior unchanged.
Kristian Bendiksen