SpMM-Bench: Performance Characterization of Sparse Formats for Sparse-Dense Matrix Multiplication

Sparse-Dense Matrix Multiplication (SpMM) plays a key role in fields like machine learning, scientific simulations, and graph analytics. However, compared to Sparse Matrix-Vector Multiplication (SpMV) and Sparse-Sparse Matrix Multiplication (SpGEMM), SpMM has received relatively less attention. To address this gap, we introduce SpMM-Bench, a benchmark suite designed to evaluate SpMM performance across various sparse matrix formats—such as COO, CSR, ELLPACK, and BCSR—on heterogeneous hardware platforms, including Arm and x86 architectures, as well as GPUs.SpMM-Bench provides a flexible framework for analyzing how factors like sparsity patterns, matrix sizes, density distributions, and hardware-specific features impact computational performance. Our suite includes parallel implementations of SpMM kernels using OpenMP for multicore CPUs and GPUs, ensuring compatibility with modern parallel computing paradigms. Through extensive evaluations, we uncover key performance trade-offs influenced by thread scaling, k-loop bounds, and block size configurations. For example, on Arm platforms, BCSR and ELLPACK formats excel with high thread counts, while on x86 architectures, CSR and COO formats perform better under certain matrix characteristics. GPU results highlight the benefits of hardware acceleration. We compare our implementations with NVIDIA’s cuSparse library, revealing opportunities to enhance GPU performance. Manual optimizations like loop unrolling and SIMD utilization significantly improve performance on x86 CPUs. These findings guide matrix format selection, parameter tuning, and hybrid execution strategies, addressing key gaps in sparse linear algebra research. The source code and test results of SpMM-Bench are released as BSD-licensed open-source from https://github.com/passlab/spmm-bench.