Fine-Grain GPU Parallelization of the Generalized Partition Crossover for Large-Scale Traveling Salesman Problems
Read the original on arXiv AI →The paper introduces a fine‑grain GPU implementation of the partition phase of the Generalized Partition Crossover (GPX) for large‑scale Traveling Salesman Problem (TSP) instances. By reformulating GPX partitioning as a graph‑parallel problem with coalesced memory layouts, ghost‑node transformations, and connected‑component analysis, the authors parallelize key operations such as union of parent tours, splitting of degree‑four vertices, deletion of common edges, and component identification using CUDA. Experiments on instances from 10,000 to 2 million cities show speedups between 48× and 625× over a naive sequential CPU implementation while significantly reducing memory overhead.
Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.