发表机构
Hefei University of Technology, Hefei, China; Lingyang Industrial Internet Co., Ltd., Hefei, China; Kochi University of Technology, Kochi, Japan（合肥工业大学; 灵阳工业互联网有限公司; Kochi大学技术）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
提出WBMM方法，通过窗口化批量矩阵乘法替代逐点卷积，实现大感受野下的高效计算，在ImageNet等任务上取得加速与精度提升。
AI 中文摘要
大核深度可分离卷积性能优异，但随着核尺寸增大，基于收集计算的不规则内存访问导致性能严重下降；而大核加速（LKA）在小特征图上有效，在大特征图上适得其反，甚至比非加速实现更慢。我们提出窗口化批量矩阵乘法（WBMM），将输入划分为连续窗口，并索引紧凑的相对位置偏置表来构建权重矩阵，通过批量矩阵乘法实现规则内存访问。这产生了一个独特性质：WBMM的吞吐量随窗口增大而提高，与深度可分离卷积随核增大而性能下降相反。算子级基准测试显示，14x14窗口的WBMM在速度上优于5x5深度可分离卷积基线，同时提供每层7.8倍更大的感受野。结合块间跨窗口通信和分层窗口重参数化，WBMM在ImageNet-1K、COCO和ADE20K上达到相当或更高的精度，训练加速1.31-1.88倍，并在GPU、CPU和边缘设备上展示出一致优势，无需专用加速内核。我们的代码可从此网址获取。
英文摘要
Large kernel depthwise convolutions achieve strong performance but suffer from significant degradation as kernel size grows due to irregular memory access from gather-based computation; while Large Kernel Acceleration (LKA) helps on small feature maps, it becomes counterproductive on large feature maps, even slower than non-accelerated implementations. We propose Windowed Batch Matrix Multiplication (WBMM), which partitions input into contiguous windows and indexes a compact relative position bias table to construct weight matrices, enabling regular memory access via batched matrix multiplication. This yields a unique property: WBMM's throughput improves with larger windows, opposite to depthwise convolutions that degrade with larger kernels. Operator-level benchmarks show WBMM with 14x14 windows outperforms 5x5 depthwise convolution baselines in speed while providing a 7.8x larger per-layer receptive field. Combined with inter-block cross-window communication and hierarchical window reparameterization, WBMM achieves comparable or higher accuracy on ImageNet-1K, COCO, and ADE20K with 1.31-1.88x training speedup, and demonstrates consistent advantages across GPU, CPU, and edge devices without requiring specialized acceleration kernels. Our code is available at https://github.com/wansong-s/WBMM.
Comments23 pages, 4 figures. Accepted as a Spotlight paper at ICML 2026. Code available at https://github.com/wansong-s/WBMM
