MatXtract: Sparsity-Aware Matrix Transformation via Cascaded Compute Density EXtraction for SpMV
Open Access
- 22 March 2026
- journal article
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (1) , 1-24
Abstract
Sparse Matrix-Vector multiplication (SpMV) is a fundamental kernel across scientific computing and graph analytics. Modern GPUs feature specialized processing units such as Tensor Core Units (TCUs) for accelerating dense matrix operations. While recent efforts leverage TCUs for SpMV acceleration, direct TCU-only mappings suffer from poor cache utilization and limited performance generalization across diverse sparse matrices. This article presents MatXtract, 1 a sparsity-aware matrix transformation framework that restructures sparse matrices through cascaded compute density extraction, enabling efficient TCU-accelerated SpMV. MatXtract introduces three key components: (1) Compute Density Extraction optimizes memory hierarchy utilization while creating Tensor Core- and CUDA core-friendly computation substrates to enhance TCU compute density and CUDA core load balancing; (2) Cascaded Execution Kernel employs a two-stage cascaded compression that systematically exploits hierarchical sparsity to maximize GPU computational throughput; and (3) Sparsity-Aware Predictor mitigates the mismatch between fixed TCU computation and diverse sparsity through Bayesian optimization for enhanced performance generalization. Extensive evaluations on 2,059 matrices show that MatXtract outperforms cuSPARSE, DASP, CSR5, and Merge-SpMV, achieving higher performance on 96.64% of matrices, with an average 1.98× and up to 8.83× speedup over cuSPARSE on NVIDIA A100 GPUs.Keywords
Funding Information
- National Natural Science Foundation of China (62372432)
- National Key Research and Development Program of China (2023YFB3001700)
This publication has 72 references indexed in Scilit:
- HotTiles: Accelerating SpMM with Heterogeneous Accelerator Architectures2024
- Variable-Sized Blocks for Locality-Aware SpMV2021
- Unleashing the Low-Precision Computation Potential of Tensor Cores on GPUs2021
- FinePar: Irregularity-aware fine-grained workload partitioning on integrated architectures2017
- LightSpMV: Faster CSR-based sparse matrix-vector multiplication on CUDA-enabled GPUs2015
- CUDA-enabled Sparse Matrix–Vector Multiplication on GPUs using atomic operationsParallel Computing, 2013
- AdELL: An Adaptive Warp-Balancing ELL Format for Efficient Sparse Matrix-Vector Multiplication on GPUs2013
- A new approach for sparse matrix vector product on NVIDIA GPUsConcurrency and Computation: Practice and Experience, 2010
- Automatically Tuning Sparse Matrix-Vector Multiplication for GPU Architectures2010
- Hypergraph-partitioning-based decomposition for parallel sparse-matrix vector multiplicationIEEE Transactions on Parallel and Distributed Systems, 1999