ENH: Use one fast exact duplicate-vertex finder in all duplicate-vertex code

Ouverte
#1,920 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 6 jours

Personne n'a encore pris cette issue.

Évaluation

Difficulté
5/5
Temps estimé
Plus d'une semaine
Accessibilité débutants
35/100
Type d'issue
Fonctionnalité
Clarté
Plutôt claire
Activité
Active
Stack technique
cpp

Piste de recherche

Start by reading GeometryUtilities::EliminateDuplicateNodes in src/simplnx/Utilities/GeometryUtilities.cpp and MeshingUtilities::OrderSharedVertices / HasDuplicateVertices in src/simplnx/Utilities/Meshing/VertexUtilities.cpp, along with issue #1908 item L19. The issue leaves the algorithm and memory budget open, so first compare the existing paths against the exact-equality, parallel, cancellable, out-of-core requirements. Done means one shared finder is used by all listed filters and keeps the first vertex in each duplicate group.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

enhancement Out-of-Core

Summary

Request: one shared, fast duplicate-vertex finder that every filter which looks for duplicate vertices uses. It should find vertices whose X, Y and Z coordinates are exactly equal, run in parallel, scale with the number of vertices, and work with out-of-core (OOC) data.

Why it is needed

simplnx has two separate duplicate-vertex finders. They have different weaknesses:

  • GeometryUtilities::EliminateDuplicateNodes (GeometryUtilities.hpp:119, FindUniqueIdsImpl) puts each vertex into a spatial bin and compares, in parallel, every pair of vertices inside each bin with exact ==. It is fast for typical STL files, but:
    • The grid always has 100 × 100 × 100 bins, whatever the number of vertices. The comparison inside a bin is O(k²). A flat mesh uses only one layer of 100 × 100 bins, and dense lattice meshes (surface meshes made from voxel data) put many vertices into each bin.
    • The bins (one std::vector per bin, about 24 MB when empty) and the Int64DataStore of unique IDs are always in memory, so the memory use is about 16 bytes or more per vertex, with no OOC path.
    • A NaN coordinate makes the bin index cast undefined behavior, and an infinite coordinate breaks the bounding box. Neither case is checked.
  • MeshingUtilities::OrderSharedVertices / HasDuplicateVertices (VertexUtilities.cpp) sort the vertex indices and compare neighbors. Issue #1908 item L19 (same defect as #1783) changes this code to exact equality with a sort on (x, y, z, input index). That sort is O(N log N), single threaded, and reads each coordinate through the DataStore's virtual accessor. With an OOC store, these reads come in random order through the chunk cache, because the sort compares vertices in random order.

Filters affected

  • Through EliminateDuplicateNodes: Read STL File, Combine STL Files (it runs Read STL File for each file), Slice Triangle Geometry and Extract Feature Boundaries 2D.
  • Through OrderSharedVertices / HasDuplicateVertices: Identify Duplicate Vertices and Verify Triangle Winding.

Expected behavior (the rule chosen for L19)

  • Two vertices are duplicates only when X, Y and Z are exactly equal (==, so -0.0 equals +0.0). A vertex with a NaN coordinate is never a duplicate.
  • In each group of equal vertices, the first vertex in input order is kept.
  • The run can be cancelled.

Measurements (from the L19 review, macOS ARM64, in memory, one thread, about 4 million vertices on a 200³ lattice in random order, no duplicates): the old one-axis sort took 0.42 s (but it misses duplicates), and the L19 sort takes 1.30 s. A scan that compares all vertices with the same primary coordinate took about 117 s.

Still open: the algorithm (hash on the exact coordinate bits, bins sized from the vertex count, or a sort that runs on chunks and then merges them for OOC data) and the memory budget that is allowed.

Langage dominant
C++
Étoiles
18
Forks
15
Merge moyen
4 j 23 h
PR mergées (30 j)
42

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de BlueQuartzSoftware/simplnx

Toutes les issues de BlueQuartzSoftware/simplnx

Issues similaires

Plus d'issues C++

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.