The reason for doing this now is that @imadmali is putting a bunch of new models into rstanarm. For one class of models (ICAR models) it can be done fairly efficiently in the Stan language (using a method due to @mitzimorris), but for models with higher-order interactions, this is really inconvenient, and it’s much easier to just have a quadratic form v’Qv for a (data) sparse matrix Q. Imad tried implementing this using the existing csr_matrix_times_vector function but that was twice as slow as Mitzi’s version, which uses a well-deployed dot_self.
There is an ongoing discussion on here Sparse matrix use cases (resolving declaration issues) about how to acutally build that interface, but it hasn’t finished yet.