Open ocramz opened 7 years ago
I have already some parts of a CSR and CSB (compressed sparse blocks) implementation
The benefits are unclear at this stage but many algorithms which require exclusively mat-vec (but not mat-transpose-vec) such as BiCGSTAB might benefit from it
This also benefits from having typeclass-based algorithms, see #28
The QR factorization is extremely slow with current IntM (IntMap-of-IntMap) based backend, which means that GMRES is slow, which is not good.
IntM
I have already some parts of a CSR and CSB (compressed sparse blocks) implementation
The benefits are unclear at this stage but many algorithms which require exclusively mat-vec (but not mat-transpose-vec) such as BiCGSTAB might benefit from it