Use Preferences to toggle precompilation of code that requires GPL dependencies

[DilumAluthge]

Fixes #224

We use Preferences. The user can toggle the behavior by specifying the include_sparse preference. If the user does not specify the preference, we default to the value of Base.USE_GPL_LIBS.

[codecov]

Codecov Report

Merging #238 (fb6fd69) into main (0857619) will decrease coverage by 0.18%.
The diff coverage is 62.50%.

@@            Coverage Diff             @@
##             main     #238      +/-   ##
==========================================
- Coverage   65.04%   64.86%   -0.19%     
==========================================
  Files          11       12       +1     
  Lines         701      703       +2     
==========================================
  Hits          456      456              
- Misses        245      247       +2     

Impacted Files	Coverage Δ
src/LinearSolve.jl	75.00% <ø> (ø)
src/factorization.jl	78.43% <ø> (-0.21%)	⬇️
src/default.jl	46.47% <50.00%> (-1.35%)	⬇️
src/factorization_sparse.jl	100.00% <100.00%> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[andreasnoack]

Ref #224

[ChrisRackauckas]

What about KLUFactorization and UMFPACKFactorization? Won't those need to be avoided? Then won't the default algorithm choice need to pick some other algorithm for sparse? What algorithm would it choose, does Julia have a fallback here?

[DilumAluthge]

In the case where Julia was built with USE_GPL_LIBS=0, I think it's fine for us to say that this package doesn't support any functionality on sparse arrays.

[rayegun]

Nothing in SparseArrays for sure.

Only factorizations are GPLd though?

[ChrisRackauckas]

So then what do we do with the default algorithm choice?

[DilumAluthge]

We could check if LinearSolve.INCLUDE_SPARSE is false, and if it is false, if A is a sparse matrix, then defaultalg(A, ...) throws an error saying that sparse matrices are not supported?

[ChrisRackauckas]

But then how do we make that dispatch on A?

[DilumAluthge]

The SparseMatrixCSC type still exists in Julia even if Base.USE_GPL_LIBS is false. So we can still dispatch on it.

[ChrisRackauckas]

And does a generic fallback implementation exist on \?

[ChrisRackauckas]

@Wimmerer weren't you going to do a pure Julia KLU sometime?

[ChrisRackauckas]

I guess we can default it to KrylovJL_GMRES

[j-fu]

And does a generic fallback implementation exist on \?

https://github.com/PetrKryslUCSD/Sparspak.jl

[DilumAluthge]

And does a generic fallback implementation exist on \?

Yeah:

julia> Base.USE_GPL_LIBS
false

julia> using SuiteSparse, SparseArrays, Random, LinearAlgebra

julia> A = sprand(4, 4, 0.3) + I
4×4 SparseMatrixCSC{Float64, Int64} with 6 stored entries:
 1.24017    ⋅    ⋅         ⋅
  ⋅        1.0   ⋅         ⋅
 0.984142   ⋅   1.0        ⋅
  ⋅         ⋅   0.824643  1.0

julia> b = rand(4)
4-element Vector{Float64}:
 0.9655062442597853
 0.15175107617153183
 0.004248022435501819
 0.4659030670094594

julia> typeof(A)
SparseMatrixCSC{Float64, Int64}

julia> typeof(b)
Vector{Float64} (alias for Array{Float64, 1})

julia> A\b
4-element Vector{Float64}:
  0.7785278622862858
  0.15175107617153183
 -0.7619340985421428
  1.0942264516593845

[rayegun]

It still has to be LGPL, but yeah I'll get it out the door this week @ChrisRackauckas. It just defaults to the dense case as far as I can tell. Once glue deps arrive I can add stuff like Sparspak to LinearSolve? Or I can just go ahead and do that as a direct dep.

[DilumAluthge]

I guess we can default it to KrylovJL_GMRES

Alright, I just pushed b576e40 - if INCLUDE_SPARSE is false and A is a SparseMatrixCSC, we default to KrylovJL_GMRES().

[ChrisRackauckas]

This doesn't seem right?

[DilumAluthge]

Ooops.

[DilumAluthge]

Okay I think I've reverted that change.

[ChrisRackauckas]

Once glue deps arrive I can add stuff like Sparspak to LinearSolve? Or I can just go ahead and do that as a direct dep.

I think it's fine to add it as a direct dep

[j-fu]

Once glue deps arrive I can add stuff like Sparspak to LinearSolve? Or I can just go ahead and do that as a direct dep.

I think it's fine to add it as a direct dep

@PetrKryslUCSD