Authors: Kaige Zhang (Beihang University); Tianyu Feng (Beihang University; Hong Kong University of Science and Technology, Guangzhou); and Xin You, Da Huo, Kejie Ma, Kelun Lei, Zhongzhi Luan, Yi Liu, Hailong Yang, and Depei Qian (Beihang University)
Abstract: Sparse matrix-dense matrix multiplication (SpMM) is critical for graph neural networks, yet GPU optimization faces challenges from irregular sparsity and thread-block load imbalance. Existing hand-tuned solutions tightly couple format design with kernel implementation, requiring reimplementation when the format changes. We present Sparsene, a format-driven automated framework that derives all kernel-level optimizations from a single sparse format specification. A hierarchical DSL decouples format structure from implementation via declarative blocking primitives and dynamic lowering transformations. The frontend compiles the DSL into a pGraph of pTasks, whose granularity and dependencies are determined by the format. This format-derived pGraph drives both automatic pipeline construction, where a hardware-aware simulator and two-stage hybrid search identify high-quality pipeline plans, and format-aware load balancing, which selects among binding strategies through per-input profiling to mitigate thread-block load imbalance. Experiments on NVIDIA A100, H100, and RTX 4090 demonstrate up to 8x speedup over state-of-the-art solutions with performance portability.
Presentation: file