Preallocate workspace for solvers about krylov.jl

[amontoison]

Speak of the devil ! We were talking about it this afternoon with Geoffroy.

[amontoison]

I have something in mind, I will open a pull request when it's ready. It should not break GPU support, which a little bit tricky.

[amontoison]

@geoffroyleconte
I let you try to add this feature for MINRES.
To not break GPU support you should parameterize your solver :

abstract type KrylovSolver{S} end
mutable struct MinresSolver{S} <: KrylovSolver{S}
  x  :: S
  r1 :: S
  ...
end

where S is Vector{Float64}, CuVector{Float64}, ... It should not be Float32, Float64, ...
You can find the variable names of MinresSolver here : https://github.com/JuliaSmoothOptimizers/Krylov.jl/blob/master/test/test_alloc.jl#L41

Like Dominique proposed you can create a function like that after :

function minres(A, b :: AbstractVector{T}; kwargs...) where T <: AbstractFloat
  solver = ... # create your MinresSolver 
  minres!(A, b, solver; kwargs...)
end

In minres!, you can add the following lines at the beginning to reuse all the actual code of minres :

x = solver.x
r1 = solver.r1
....

You only need to update the initialization of the variables and add a check to verify that eltype(S) == T.

[dpo]

😈

[amontoison]

fixed by #271