GSPARC: A Mode-Aware GPU-Based Method for Large-Scale Sparse Tensor Contraction

Abstract

Sparse tensor contraction (SpTC) is a core operation for high-order sparse data analytics in recommendation, knowledge graph analytics, large-scale tensor mining, and scientific computing. Existing methods are limited by SpTC-misaligned representations, overlapping outputs across partitions, and memory-inefficient accumulation over large output spaces. We propose GSPARC, a GPU-based SpTC method that combines a new mode-aware bit-wise representation with mode-decoupled output partitioning and memory-efficient accumulation. These techniques improve index matching, eliminate inter-partition merging, and avoid full output-space allocation. Extensive experiments demonstrate that GSPARC outperforms state-of-the-art SpTC methods in contraction speed and scalability. Our code is available at this link: https://github.com/seohyeon0322/GSPARC.