Open perazz opened 10 months ago
Hi,
In order to help this idea move forward regarding its extension to sparse algebra I thought about moving parts of FSPARSE here.
I thought about doing a first PR to define the sparse types with something like:
Other formats can be added following a similar pattern. Before going any further with the methods, I though that having some comments about the derived types for the data containment is important.
For instance:
@jalvesz it seems a good start to me. I developed a library with similar interfaces. There have been already a lot on discussion about sparse matrices, and it is always difficult to find a clear clonclusion. So, I suggest to keep the first draft a simple as possible. Other formats (and dense matrices) could be added later.
Would it make sense to open a new branch on the stdlib repo to coalesce linear algebra progress around?
Would it make sense to open a new branch on the stdlib repo to coalesce linear algebra progress around?
Yes, it is probably a good idea. There is also #189 related to this topic that is open since a while (cc @jalvesz )
There is also this page related to sparse matrices
Let me know when you open the branch and I'll move the PR under that one, indeed it would be easier to coalesce all linear algebra topics under one branch that could move a bit faster and serve as playground
Hey @perazz is this related to the grant money from STF and the work done over at https://github.com/perazz/fortran-lapack or would this be an independent piece of work (to be performed by the community)? Regardless, it would make sense to have an stdlib branch containing the relevant progress.
Hello stdlib developers, I'm opening this issue to summarize and coalesce upcoming efforts to integrate linear algebra operations in stdlib, in particular:
1) Accessible interfaces for common linear algebra operations
We want stdlib to be able to solve linear algebra tasks by wrapping against libraries like BLAS, LAPACK, SCALAPACK, with a user-friendly interface. I think this means:
2) Support for common IO formats for matrices and tensors
To allow efficient working with multidimensional arrays, we should support easy serialization/deserialization and conversion among formats. This could include formats like NPZ, matrixmarket, etc. An easily extendible API should be provided to have plugins for other formats which might or might not fit in the scope of stdlib.
I believe an equally important task is to define derived types that are capable of storing temporary information which is not strictly matrix data (e.g. matrix factorization, or working arrays for the matrix solvers), to avoid unnecessary overhead in case of repeated algebra operations. This means that "simple" array storage may need to be replaced with matrix derived types in those cases.
Because there are plenty of options in defining these APIs, It is crucial to the success of this task that as much feedback as possible is given, so I would like to encourage all ideas - and criticisms - to be discussed on this issue, so that we can come up with the best possible version. I am also opening a discussion page on the Fortran-lang Discourse that we can use for more verbose discussions.
Thank you, Federico
cc @certik @awvwgk @fortran-lang/stdlib @fortran-lang/admins
Linked issues
Linear algebra (BLAS/LAPACK)
1
10
67
450
476
-> Regarding dense algebra, I've started a discussion at https://github.com/fortran-lang/stdlib/issues/450
Sparse algebra
38