Core Concepts
GPolylla introduces a novel GPU-accelerated algorithm for generating polygonal meshes, achieving significant speedups compared to CPU implementations.
Abstract
The article presents GPolylla, a fully GPU-accelerated polygonal mesh generator. It discusses the importance of polygonal mesh generation in numerical methods like the Virtual Element Method (VEM) and compares the advantages of Polylla over existing algorithms. The work details the design and implementation of GPolylla, emphasizing its use of a half-edge data structure on GPUs for efficient mesh topology management. Experimental results demonstrate substantial speedups compared to CPU implementations.
-
Introduction
- Importance of polygonal mesh generation in numerical methods.
- Advantages of Polylla for generating arbitrary polygonal shapes.
-
Meshing Concepts
- Introduction to longest-edge propagation path, terminal-edge regions, and frontier-edges.
-
Polylla Algorithm
- Phases: Label, Traversal, Repair.
-
Mesh Representation: CPU and GPU
- Use of half-edge data structure for efficient mesh processing.
-
Secuential Polylla
- Description of the sequential version's label phase, traversal phase, and repair phase.
-
GPU Polylla
- Explanation of GPU-accelerated algorithms including longest edge labeling, frontier-edges labeling, seed-edges labeling, attribute changes, search for frontier edges for each seed edge, overwriting seeds.
-
Experiments
- Dataset description: Grid meshes and Random meshes.
- Experimental setup on Patagón supercomputer with Nvidia DGX A100 GPU node.
-
Contribution
- Introduction to new version of sequential Polylla algorithm with output as half-edge data structure.
Stats
The experimental results show up to ×83.2 speedup compared to CPU sequential implementation.
Speedup reaches ×746.8 when excluding data structure copying costs.