IAP-SpTV: An input-aware adaptive pipeline SpTV via GCN on CPU-GPU

Abstract

Sparse tensor-times-vector (SpTV) is the core computation of tensor decomposition. Optimizing the computational performance of SpTV on CPU-GPU becomes a challenge due to the complexity of the non-zero element sparse distribution of the tensor. To solve this problem, we propose IAP-SpTV, an input-aware adaptive pipeline SpTV via Graph Convolutional Network (GCN) on CPU-GPU. We first design the hybrid tensor format (HTF) and explore the challenges of the HTF-based Pipeline SpTV algorithm. Second, we construct Slice-GCN to overcome the challenge of selecting a suitable format for each slice of HTF. Third, we construct an IAP-SpTV performance model for pipelining to achieve the maximum overlap between transfer and computation time during pipelining. Finally, we conduct experiments on two CPU-GPU platforms of different architectures to verify the correctness, effectiveness, and portability of IAP-SpTV. Overall, IAP-SpTV provides a significant performance improvement of about 24.85% to 58.42% compared to the state-of-the-art method.