GPU-Accelerated Graph Partitioning Algorithms in VLSI Design
Guardat en:
| Publicat a: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Autor principal: | |
| Publicat: |
ProQuest Dissertations & Theses
|
| Matèries: | |
| Accés en línia: | Citation/Abstract Full Text - PDF |
| Etiquetes: |
Sense etiquetes, Sigues el primer a etiquetar aquest registre!
|
| Resum: | Graph and hypergraph play critical roles in computer‑aided design (CAD) because it allows us to break down a large circuit into several manageable pieces that facilitate efficient CAD algorithm designs. However, as circuit sizes continue to grow, partitioning becomes increasingly time‑consuming. To address this runtime bottleneck, many researchers have leveraged CPU-parallel techniques to accelerate partitioning. However, the speedups are typically limited to 8–16 CPU threads. To overcome this challenge, this thesis leverages the massive parallelism of GPUs and introduces two GPU-accelerated partitioning algorithms: iG-kway for graphs and iHyperG for hypergraphs. G-kway features a union find-based coarsening algorithm that merges many vertices simultaneously, and a novel independent-set-based refinement algorithm that refines thousands of vertices in parallel. HyperG introduces a balanced group coarsening method and a sequence-based refinement algorithm. Experimental results show that G-kway achieves an average speedup of 8.6x over the CPU-parallel graph partitioner mt-metis, while HyperG delivers an average speedup of 4.1x over the CPU-parallel hypergraph partitioner Mt-KaHyPar, both maintaining comparable cut quality. While G-kway and HyperG achieve new performance milestones in graph and hypergraph partitioning, they are limited to full partitioning. This lack of support for incrementality presents a critical limitation for many CAD applications, where circuits undergo iterative modifications as part of optimization loops. To address this limitation, this thesis also introduces two GPU-parallel incremental partitioning algorithms: iG-kway for graphs and iHyperG for hypergraphs. iG-kway features an incrementality-aware bucket-list data structure and a refinement kernel that refines only affected vertices. iHyperG introduces a scalable delta-based hypergraph data structure for efficient incremental modifications on the GPU, along with an effective incremental partitioning algorithm that rebalances the partition in a single pass and refines only cut-critical vertices. Experimental results show that iG-kway achieves an average speedup of 84x over the GPU-based full graph partitioner G-kway, while iHyperG delivers 190x speedup for hypergraph modification and 83x for partitioning over the state-of-the-art GPU-based full hypergraph partitioner, both maintaining comparable cut quality. |
|---|---|
| ISBN: | 9798270241681 |
| Font: | ProQuest Dissertations & Theses Global |