Delaunay Mesh Refinement on the GPU

Zhenghai Chen       Tiow-Seng Tan       Hong-Yang Ong

School of Computing
National University of Singapore

{dcschai | dcstants} | dcsohy} @ nus.edu.sg

Abstract

We present three GPU algorithms for performing 2D and 3D Delaunay mesh refinement.  For 2D, we present a GPU constrained Delaunay algorithm which takes in a plane straight line graph as input and generates a conforming quality mesh.  For 3D, we present a GPU constrained Delaunay algorithm and a GPU restricted Delaunay algorithm, which both take in a piecewise linear complex as input and generates a conforming quality mesh and an approximating quality mesh respectively.  These algorithms adopt a set of design strategies, aimed at improving GPU utility, which can be possibly applied to other mesh refinement solutions. Experimental results show that our algorithms are faster than the current state-of-the-art counterparts (for both sequential and multi-threading CPU versions) by up to an order of magnitude, while maintaining similar number of Steiner points being inserted. This may make Delaunay mesh refinements feasible in interactive applications. 

 

Keywords: Parallel Meshing, Delaunay Triangulation, Little¡¯s Law

 

Manuscript (15 pages) ¨C Some experimental results can be found in Chen¡¯s Ph.D thesis (PPT for the Ph.D defence)

 

Software Download:

¡¤         gDP2d: 2D Constrained Delaunay Refinement

o   zip file: source and executable + data, 394MB, or

o   zip file: executable + data, 380MB (for windows), or

o   see github

o   Comparison is done with Triangle

¡¤         gQM3d: 3D Constrained Delaunay Refinement

o   zip file: source and executable + data, 87MB, or

o   zip file: executable + data, 45MB (for windows), or

o   see github

o   Comparison is done with Tetgen and CGAL

o   May use Medit to view the output

¡¤         gDP3d: 3D Restricted Delaunay Refinement

o   zip file: source and executable + data, 59MB, or

o   zip file: executable + data, 56MB (for windows), or

o   see github

o   Comparison is done with CGAL

o   May use MeshLab / Medit to view the output

Sample outputs:

(a)   gDP2d: 2D Constrained Delaunay Refinement

(i)          Quality Mesh from a raster image

(ii)         Synthetic Data

gdp2D_4distributions

(b)  gQM3d: 3D Constrained Delaunay Refinement

(i)          Quality Mesh of an elephant model

gQM3d_banner

(ii)         The distributions of dihedral angles in the output triangulations of the Sculpture model (65942, in Thingi10K)

gQM3d_dihedral_cutoff

 

(c)   gDP3d: 3D Restricted Delaunay Refinement (Voronoi Lamp, Model 940414, in Thingi10k)

gdp3D_banner

 

Other triangulation projects with GPU: Delaunay Triangulation, Constrained Delaunay Triangulation, 3D Delaunay Triangulation,

 

Updated Dates: 21 September 2020, 24 April 2021

 

© School of Computing, National University of Singapore