Quasi-Newton: Use LAPACK for LU factorization

[MaxenceGollier]

Penelopt.jl/src/linear_algebra/ldlt.jl

Line 343 in 38e5313

F = lu!(Z2[1:(2*p), 1:(2*p)], check = false) # FIXME ?

We can make this allocation free, for example, Claude gave

using LinearAlgebra

n = 8
A = rand(n, n)
ipiv = Vector{LinearAlgebra.BlasInt}(undef, n)  # preallocate once

# In your hot loop:
LinearAlgebra.LAPACK.getrf!(A, ipiv)   # in-place LU, no allocation

[amontoison]

@MaxenceGollier I suggest to directly call the Fortran routine from Julia.
I did a few Julia wrappers in MadNLP.jl and you can reuse the buffer this way: https://github.com/madsuite-org/MadNLP.jl/blob/master/src/LinearSolvers/lapack.jl#L50

A bunch of routines in LinearAlgebra.LAPACK.*! are in in-place in terms of operations but still allocate (LBL', QR, EVD, etc...).
They don't cache the buffer and call the routine twice (one to get the buffer size and the second one to perform the operation).

Cholesky and LU don't need a buffer (on CPU).

[MaxenceGollier]

Even when i don't request your help, you appear out of nowhere to fix my issues... What a legend.

Thank you @amontoison.

[amontoison]

You appeared in my GitHub news yesterday evening because I follow you.
I got 2 hours to code on something unrelated to work. Nice coincidence 🙂