compact dense sparsity patterns about casadi HOT 3 OPEN

The sparsity pattern is stored as a vector of integers where:

• The first two elements are the dimensions (nrow, ncol)
• The following (ncol+1) entries are the nonzero offset for each column. Note that this means that the (ncol+3)-th entry is the total number of nonzeros (nnz)
• The last nnz entries are the row index for each nonzero. Indices are increasing for each column, with no duplicates

For a dense matrix, we have nnz=nrow * ncol, which can easily be checked, and the last part is superfluous information. So you can decrease the storage needs from (2+(ncol+1)+nrow * ncol) to (2+(ncol+1)). The problem is that every algorithm where sparsity is used needs to be specialized to handle the dense case separately. Currently, many common algorithms are, like matrix multiplication, but not all. For maintainability reasons, I would be against implementing the more compact storage (which is easy and quick to do) unless we go through all algorithms in CasADi where sparsity is being used and update them accordingly. You can of course implement it yourself in your own fork of CasADi.

It's possible to get an even more compact representation, or if you want to support something like triangular patterns or higher order tensors, by letting the third element of the sparsity vector be something else than zero. It's always zero now by definition. But again, this generation would need to be propagated everywhere in a maintainable way.

from casadi.

Okay, thanks for the feedback! I'll leave this open for now since it seems like a valid feature.

from casadi.

Okay, thanks for the feedback! I'll leave this open for now since it seems like a valid feature.

It would be a useful feature. It's usually hard to get funding for features that would "benefit a lot of users a little" as opposed to "benefit a few users a lot". That's the situation with a lot of potential core improvements.

from casadi.