Simple rules for optimal sparse-format selection

Identify a set of simple matrix-profile rules that can select the optimal sparse matrix storage format for sparse matrix–vector multiplication without requiring a complex machine-learning procedure.

Background

The paper evaluates machine-learning models that select among nine sparse matrix storage formats for SpMV on SpacemiT K1 and K3 RISC-V processors. Although the models achieve substantial speedups over scalar CSR, their predictions remain sensitive to the training data and are not consistently optimal. Feature-importance analysis also shows that the relationships among matrix characteristics are complex and vary with processor architecture and numerical precision.

The authors therefore leave unresolved whether the relevant matrix-profile characteristics can be reduced to a transparent set of simple selection rules. Solving this problem would make format selection easier to interpret and deploy at SpMV call sites, while avoiding the complexity and instability of the current machine-learning-based procedure.