AbstractSparseMatrixCSC not exported

[j-fu]

Hi, I would like to use the AbstractSparseMatrixCSC interface with my sparse matrix struct, however it seems that it is not exported.

Also, it is not completely clear what needs to be implemented to conform to the interface.
I currently assume it is getcolptr, getrowval, nonzeros, nnz, size.

[rayegun]

That should be sufficient for most operations yes. There will definitely be some things missing, there are not many other subtypes of AbstractSparseMatrixCSC in the ecosystem to test it out.

You can qualify the name SparseArrays.AbstractSparseMatrixCSC, but I'll look into adding it to exports.

[j-fu]

... bump...

[ViralBShah]

Is it possible to provide a PR? We should also document the interface.

[j-fu]

Ok will try. Can get on this in a couple of days.

[j-fu]

What the sparse(A) method should do ?
Currently, we have

SparseArrays.jl/src/sparsematrix.jl

Line 1000 in 0d2d30a

sparse(S::AbstractSparseMatrixCSC) = copy(S)

and

SparseArrays.jl/src/sparsematrix.jl

Line 729 in 0d2d30a

SparseMatrixCSC(S::AbstractSparseMatrixCSC) = copy(S)

(and following).

All of them copy the AbstractSparseMatrixCSC. IMHO all of them could return a SparseMatrixCSC via this constructor,
and I think the wordl would be a better place if they would.

So the question here is: is a sparse(A::AbstractSparseMatrixCSC)::SparseMatrixCSC method to be seen as a part of the AbstractSparseMatrixCSC interface ?

[ViralBShah]

If you have your own concrete implementation of AbstractSparseMatrixCSC, then sparse should give you a copy of the matrix in that concrete type, I believe. That way, I believe, your type will propagate as much as possible through the implementation.

[j-fu]

Ok, that may be the philosophy behind sparse. But then why currently some of the constructors of SparseMatricCSC from AbstractSparseMatrixCSC do copy too ?

[ViralBShah]

The AbstractSparseMatrixCSC was not in the original design, but introduced later on by folks who want to roll out their own. So it was a mechanism for hooking in by whoever needed it last and needs to keep improving.

[j-fu]

... this is exactly my reason to try to help to move this forward :)

[j-fu]

But still: why do we need these ?

SparseArrays.jl/src/sparsematrix.jl

Line 729 in 0d2d30a

SparseMatrixCSC(S::AbstractSparseMatrixCSC) = copy(S)

SparseArrays.jl/src/sparsematrix.jl

Line 733 in 0d2d30a

SparseMatrixCSC{Tv,Ti}(S::AbstractSparseMatrixCSC{Tv,Ti}) where {Tv,Ti} = copy(S)

In both cases, one would expect a SparseMatrixCSC as a result and not whichever else subtype of AbstractSparseMatrixCSC...

EDIT: @tkf, @KlausC

[ViralBShah]

Yes, for the SparseMatrixCSC constructor, it should be a SparseMatrixCSC, but for sparse, it should be available to the new abstract sparse type, right?

BTW, I always felt it was a bit weird to have AbstractSparseMatrixCSC and the right thing would have been to have AbstractSparseMatrix.

[j-fu]

So a PR should fix this ?

BTW, I always felt it was a bit weird to have AbstractSparseMatrixCSC and the right thing would have been to have AbstractSparseMatrix.

... I can understand that one - there is at least the SparseMatrixCSR storage scheme which is also an AbstractSparseMatrix:

https://github.com/gridap/SparseMatricesCSR.jl/blob/d26c461eefbd27550529e06e2cafae24265657ea/src/SparseMatrixCSR.jl#L28

[KlausC]

But still: why do we need these ?

I agree with 729, 731, 733 are wrong; they should all fall back to what is defined in line 734

For line 1000 I would prefer sparse(s::Abstract...) = s. Why make a copy, it it is already sparse? If you need a copy use copy,