Dynamic Mesh Processing on the GPU (Abstract)

Ahmed H. Mahmoud, Serban D. Porumbescu, John D. Owens

Abstract: We present a parallel data structure for dynamic triangle mesh processing on entirely on the GPU. Our design supports fine-grained topological updates---e.g., element insertion, deletion, and edge flips---while maintaining high throughput and low memory overhead. The core of our approach is a topologically partitioned, update-friendly mesh representation that enables localized operations to be executed in parallel within fast on-chip memory. To manage conflicts during concurrent updates, we use a speculative processing model with lightweight rollback.Our programming model relies on the cavity operator, a general-purpose primitive for local mesh updates that replaces a neighborhood of elements in a single operation. This abstraction enables a range of dynamic algorithms---including remeshing, decimation, and refinement---to be expressed within a unified framework.We evaluate our data structure and algorithms on a suite of meshes processing applications, achieving an order of magnitude speedup over multi-threaded CPU implementations and more than two orders of magnitude speedup against single-threaded baselines. In addition, our representation outperforms static GPU mesh structures in both memory efficiency and traversal speed---demonstrating its suitability as a general-purpose substrate for parallel mesh algorithms. Our results highlight how rethinking mesh data structures around locality, parallel conflict resolution, and GPU memory hierarchy leads to significant gains for dynamic, unstructured workloads.

External IDs:dblp:conf/hopc/MahmoudPO25