中图分类号：TP303；TP183 文献标志码：A DOI: 10.16157/j.issn.0258-7998.267881 中文引用格式： 方洁虹，朱昊宇，肖宛昂. 基于算法稀疏性的卷积神经网络加速方法研究[J]. 电子技术应用，2026，52(8)：15-23. 英文引用格式： Fang Jiehong，Zhu Haoyu，Xiao Wan’ang. Research on CNN acceleration methods based on algorithmic sparsity[J]. Application of Electronic Technique，2026，52(8)：15-23.
Research on CNN acceleration methods based on algorithmic sparsity
Fang Jiehong1，2，Zhu Haoyu1，2，Xiao Wan’ang1，2
1.School of Integrated Circuits， University of Chinese Academy of Sciences；2.Artificial Intelligence and High-Speed Circuit Laboratory， Institute of Semiconductors， CAS
Abstract： While Convolutional Neural Networks (CNNs) exhibit significant sparsity after pruning and ReLU activation, which can potentially enhance inference efficiency, irregular sparsity distribution often leads to load imbalance. Furthermore, the sequential execution of convolution and pooling layers causes memory bandwidth waste, thereby limiting overall performance. To address these issues, this paper proposes a sparse CNN acceleration algorithm and A convolution-pooling fusion method, with verification conducted on the Cambricon MLU platform. First, an adaptive Sparse Matrix-Vector multiplication (SpMV) algorithm based on the Compressed Sparse Row (CSR) format, named Adaptive Sparse Convolution (AdaSpConv), is designed to effectively resolve the load imbalance among cores through dynamic task allocation. Second, a new storage format called Compressed Sparse Row-Convolution Pooling Fused (CSR-CP) is proposed to achieve operator fusion of convolution and pooling, which minimizes data movement and significantly reduces memory access latency. Experimental results demonstrate that on the MLU370-S4 platform, AdaSpConv achieves a peak performance of 0.828 GFLOP/s and an effective memory bandwidth of 3.39 GB/s, outperforming baseline Scalar， Vector， and Adaptive. Based on the CSR-CP fusion mechanism, the convolutional operations of VGG-16 and ResNet-50 are accelerated by 2.87× and 1.99×, respectively, compared to traditional methods.
Key words : Cambricon MLU；sparse matrix-vector multiplication；AdaSpConv；CSR-CP
