Yu-Hsiang M. Tsai

Yes, I am not saying we need to do switch for users but we need to provide a way such that they can switch the function. cuSPARSE only provides the...

Like a third party using ginkgo with latest feature (currently only in develop), they do not know the user ginkgo develop version. the version check can not avoid using the...

It's indeed the intermediate version without any guarantee. maybe not consider it as a version number if it's confusing with the semantic version. It's just a comparable increasing tag.

it's only based on the nonoverlapping partition. we can also add the L1 to the Schwarz for the distributed local preconditioner. 6.2 in the paper mentioned the L1 scalar Jacobi

If the partition is only designed with the mpi (let me use the mpi here to distinguish), L1 is indeed normal Jacobi i.e using Schwarz with L1 will give this...

there's one reference from precond. They use L1-scalar Jacobi but they also use the absolute value of diagonal part https://etna.math.kent.edu/vol.55.2022/pp687-705.dir/pp687-705.pdf

The first paper considers the diagonal entries are positive, so they are equivalent. Distributed on mpi or distributed on the block to me are the same thing, which are ways...

It should be theorically L1 smoother. otherwise, I just add some random Jacobi. (I need to use [] not {} to represent row index set because I can not display...

I am not sure whether to take the absolute value of the diagonal value. It might make sense for the scalar version but not for the block version. What's the...

Yes, from the block-version, it only updates the diagonal value. I think l1 is mostly from taking the summation of the absolute value in non-local block's row, not the entire...