Replace `V(undef, nvar)` by `similar(nlp.meta.x0)` in `TrunkSolver`?

[paraynaud]

Hey guys (@dpo @tmigot), I was wondering why the AbstractVector required by TrunkSolver are defined with V(undef, nvar) and not with similar(nlp.meta.x0)?

https://github.com/JuliaSmoothOptimizers/JSOSolvers.jl/blob/15550bdd2282c7fb1d7a5fcd78078ac271385b57/src/trunk.jl#L84

I'm currently working on PartitionedStructures.jl and PartiallySeparableNLPModels.jl to make solvers (starting with trunk) exploit fully the partially separable structure. I'm defining PartitionedVector<:AbstractVector in PartitionedStructure.jl, and my plan is to set nlp.meta.x0 as a partitioned vector. This way, any object similar to x0 keeps the same partitioned structure. In addition, I will overload all operations required by trunk() (and cg()) on PartitionedVectors.

Defining vectors with V(undef, nvar) prevents any use of the partitioned structure.

[dpo]

You're right. We should be using similar.

[tmigot]

Hi @paraynaud ! Could you open a PR to fix this? We now have a solver structure for all the solvers, so this issue might appear not only in TrunkSolver

[paraynaud]

I spoke with @amontoison since the same problem appears in KrylovSolvers, the way AbstractVectors are defined (with undef) seems to be the good way to do it. What I understood from Alexis, using similar in place of S(undef,nvar) will downgrade (a lot) the performances of cg if it is called with SparseVector (or an AbstractVector which is not a DenseVector).

If everyone is okay with the similar replacement I can open a PR.

Anyway, I will have to define my own TrunkSolver constructor dedicated to PartiallySeparableNLPModels, since the implementation of PartitionedVectors has become quite complex (similar is not enough unfortunately).

[amontoison]

I don't use similar in Krylov.jl because the type of the right-hand side b is not always the most appropriate to store the Krylov bases. Based on the type of b, I determine the most appropriate dense storage (S) for Krylov bases with the function ktypeof.
https://github.com/JuliaSmoothOptimizers/Krylov.jl/blob/main/src/krylov_utils.jl#L221-L248

[tmigot]

Looks good to me, I checked and only R2 uses similar right now.

[dpo]

Should we extract the Krylov mechanism out of Krylov so it can be used here too?

[amontoison]

It can be used here already :) You just need to import Krylov.ktypeof and call it.