Tenure-Track Assistant Professor at University of Pittsburgh, ECE Department; CS Ph.D.'19 UCLA
๐ฃ ๐ฃ๐ฃFCCM 2024 Program Highlight Series 07!๐ฃ ๐ฃ๐ฃ ๐๐๐-๐๐๐: ๐๐ง ๐๐๐๐ข๐๐ข๐๐ง๐ญ ๐๐๐ ๐๐๐๐๐ฅ๐๐ซ๐๐ญ๐จ๐ซ ๐๐จ๐ซ ๐๐ฎ๐ฅ๐ญ๐ข-๐๐๐ฅ๐๐ญ๐ข๐จ๐ง ๐๐๐ญ๐๐ซ๐จ๐ ๐๐ง๐๐จ๐ฎ๐ฌ ๐๐ซ๐๐ฉ๐ก ๐๐ฎ๐ญ๐ก๐จ๐ซ๐ฌ: Wenlu Peng, Jianjun Chen, Wenjin Huang, Yihua Huang ๐๐๐๐ข๐ฅ๐ข๐๐ญ๐ข๐จ๐ง: Sun Yat-sen University ๐๐๐ฒ ๐๐จ๐ง๐ญ๐ซ๐ข๐๐ฎ๐ญ๐ข๐จ๐ง๐ฌ: ๐ Data dimensionality reduction. The proposed MRH-GCN introduced a configurable multiple adjacency matrix fusion strategy designed for HGCNs, aiming to reduce the dimensionality of adjacency matrices. ๐ Data compression. This paper employed a sparse-matrix-tile space search compression algorithm to maximize the compression of sparse data in the adjacency matrix and feature matrix, reducing data sparsity and improving indexing efficiency. ๐ Regular array pipeline execution mode. The paper adopted a regular array pipeline designed for the three computation stages of HGCNs, matching memory access and computation time, and reducing on-chip memory pressure. ๐ Efficient hardware architecture. This paper employed an SpMM systolic array to facilitate the compression algorithm and regular scheduling scheme, achieving massive data parallelism. Also, relation-awareness was introduced in the combination stage to enhance computational efficiency.
