Exploring Architecture, Dataflow, and Sparsity for GCN Accelerators: A Holistic Framework
Exploring Architecture, Dataflow, and Sparsity for GCN Accelerators: A Holistic Framework
复制标题
DOI：
10.1145/3583781.3590243
复制
发表时间：
2023-06
期刊：
影响因子：
--
通讯作者：
Lingxiang Yin;J. Wang;Hao Zheng
中科院分区：
文献类型：
--
作者：
Lingxiang Yin;J. Wang;Hao Zheng
近年来，越来越多的图卷积网络（GCN）模型被应用于各种实际应用中。然而，由于图数据集的稀疏性不同，设计高效的GCN加速架构仍然具有挑战性。尽管付出了巨大的努力，但现有的工作中很少有考虑到整个GCN加速器设计的整体视图，因此，架构之间的动态交互，数据重用和并行化策略），以及压缩格式没有得到很好的研究。在本文中，我们奋进开发一个整体框架，支持快速探索GCN设计空间-协同架构，并行和稀疏优化。具体来说，我们实现了各种压缩格式，专门用于处理极端和不规则的稀疏性。此外，我们提出了一个通用的GCN架构，能够支持各种低和压缩格式在一个组合的架构。鉴于爆炸的GCN设计空间，遗传算法的实现，以促进快速探索，同时提出了一个最佳的GCN解决方案，适当的压缩和压缩格式。我们提出的框架可以实现12.3X，2.2X，和1.37倍的加速比和15.3X，3.7X，和1.6X的能源效率平均相比，HyGCN，AWB-GCN，和GCNAX，分别。
Recent years have seen an increasing number of Graph Convolutional Network (GCN) models employed in various real-world applications. However, designing efficient architectures for GCN acceleration remains challenging due to the varied sparsity across graph datasets. Despite significant efforts, very few of the existing works have considered a holistic view of the entire GCN accelerator design, and therefore, the dynamic interactions between architecture, dataflow (i.e., data reuse and parallelization strategies), and compression format are not well studied In this paper, we endeavor to develop a holistic framework supporting the rapid exploration of GCN design space - synergizing architecture, dataflow, and sparsity optimizations. Specifically, we implement a variety of compression formats tailored for handling extreme and irregular sparsity. Moreover, we propose a generic GCN architecture capable of supporting various dataflow and compression formats in one combined architecture. Given the exploded GCN design space, a genetic algorithm is implemented to facilitate rapid exploration while suggesting an optimal GCN solution with suitable dataflow and compression formats. Our proposed framework can achieve 12.3X, 2.2X, and 1.37X speedup and 15.3X, 3.7X, and 1.6X energy efficiency on average as compared to HyGCN, AWB-GCN, and GCNAX, respectively.