Authors: Luhan Wang (School of Computer Science, Peking University; Institute of Computing Technology (ICT), Chinese Academy of Sciences (CAS)); Haipeng Jia (Institute of Computing Technology (ICT), Chinese Academy of Sciences (CAS)); Kun Li (Institute for AI Industry Research, Tsinghua University); Zongyuan He (School of Computer Science, Peking University); Shengguo Li (National University of Defense Technology (NUDT)); Ting Cao and Yunxin Liu (Institute for AI Industry Research, Tsinghua University); Yunquan Zhang (Institute of Computing Technology (ICT), Chinese Academy of Sciences (CAS)); and Yifeng Chen (School of Computer Science, Peking University)
Abstract: Sparse iterative methods with static sparse matrices form the backbone of many applications, where SpMV is the primary performance bottleneck. Conventional optimizations typically involve costly matrix-specific preprocessing, causing unamortizable overhead as algorithmic advances reduce iteration counts. We present SPIKe, a practical SpMV kernel generator targeting efficient sparse iterative methods across a wide spectrum of iteration counts. SPIKe comprises three key techniques: 1) Online Thread-Level Parallelism Adaptation policy, using a thread-computed row span proxy to enhance load balance for uniform tiles; 2) Online Warp-Level Asymmetric Scheduling policy, featuring a lightweight Warp-as-a-Sensor to route skewed tiles to asymmetric code paths; 3) Offline Architecture-Domain Co-Aware Code Generation for kernel optimization following the preceding policies. Across 2,294 matrices, SPIKe outperforms six baselines with negligible preprocessing overhead. In end-to-end evaluations spanning varying iteration counts, SPIKe achieves average speedups of 2.78x for PageRank and 4.26x for GMRES against GraphBLAS and PETSc.
Presentation: file