SparseLU and CholMod offer excellent performance when repeatedly solving the same type of problem using analyze(). Is there something similar for a sparse matrix product, specifically A * S * A.transpose() where S is square and A is rectangular? Both matrices are expected to have lots of zero entries. Only non-zero entries will change in value. I'd like to repeatedly do A * S * A.transpose() as efficiently as possible.

I suppose I could write my own simple code generator and interpreter. I would only need to implement addition, multiplication, and methods to read matrix entries.

Thoughts?