SPIKe: Generating Practical SpMV Kernels for Sparse Iterative Methods on GPUs

Sparse iterative methods with static sparse matrices underpin many applications, with sparse matrix-vector multiplication (SpMV) as their primary performance bottleneck. Existing optimizations often require costly matrix-specific preprocessing whose overhead cannot be amortized as iteration counts decrease. We present SPIKe, a practical SpMV kernel generator for efficient sparse iterative methods across a wide range of iteration counts. SPIKe combines online thread-level parallelism adaptation, online warp-level asymmetric scheduling, and offline architecture-domain co-aware code generation. Across 2,294 matrices, SPIKe outperforms six baselines with negligible preprocessing overhead. In end-to-end evaluations, it achieves average speedups of 2.78x for PageRank and 4.26x for GMRES over GraphBLAS and PETSc.