Background

SpGEMM is the problem of computing the (matrix) product of two sparse matrices.

It has a few uses:

• graph algorithms, e.g. triangle counting
• machine learning, e.g. graph neural networks
• quantum chemistry

Questions

• Can I reproduce existing optimized algorithms using Finch?
• Parallel implementations of SpGEMM can have quite a few hyperparameters (see e.g. the heuristics within SuiteSparse:GraphBLAS). How can we tune these hyperparameters efficiently and automatically?

References

• A Systematic Survey of General Sparse Matrix-Matrix Multiplication
• Parallel Algorithms for Masked Sparse Matrix-Matrix Products