[General] Visite ANL

[jbcaillau]

@amontoison To be discussed:

• GPU benchmarking (allocations...) check Test with ExaModels CTBenchmarks.jl#26 (comment)
• bicoloring, coloring for ADNLP Try different order for colouring CTDirect.jl#385
• linear solvers
• kernels for matrix vector operations such $Ax$ (linear) and $\sum_i u_i B_i x$ (bilinear) with $A$ and $B$ matrices and $x$, $u$ vectors; use case: x an (1:n, 0:N) array, u an (1:m, 0:N) array for time discretisation of vector valued functions $x(t) \in \mathbf{R}^n$ and $u(t) \in \mathbf{R}^m$, and bilinear dynamics (target application: quantum control with large matrices $A$, $B_i$)

$$ \dot{x}(t) = A x(t) + \sum_i u_i(t) B_i x(t) $$

• MadNCL: to be tested soon (see benchmarking above)

[jbcaillau]

@amontoison

• enrich ExaModels grammar to allow vector constraints
• allow to parametrise an ExaModel by the grid_size (and other params...)
• allow to initialise an array variable with a constant column (or row), that is instead of this

x_l = [0, -1, 0] # lower bound for all x(t_i)
x_u = [1, 2, 0] # upper bound for all x(t_i)
ExaModels.variable(x, 3, 0:grid_size; lvar = [x_l[i] for (i, j) = Base.product(1:3, 0:grid_size)], uvar = [x_u[i] for (i, j) = Base.product(1:3, 0:grid_size)])

that builds the whole array (the copies it from CPU to GPU), write that:

ExaModels.variable(x, 3, 0:grid_size; lvar = x_l, uvar = x_u)

[jbcaillau]

@amontoison notes of the day 🙏🏽
https://github.com/control-toolbox/CTBenchmarks.jl/blob/main/kernel_test/foo.jl

[amontoison]

Reminder for May 15, 2025:

• Check together if we can pass a function f inside a kernel.

[jbcaillau]

@amontoison

• it is OK, check dynamics! (and f!) examples here
• also runs with ForwardDiff but issue somewhere (no gain), check df! in the same file
• see also: macro to generate a kernel (useless since the calling a function works)

Reminder for May 15, 2025:

• Check together if we can pass a function f inside a kernel.

[amontoison]

@jbcaillau

• Large allocations of ExaModels.jl on CPU.
• ForwardDiff.jl + GPU on pure Julia code / kernels?

[jbcaillau]

@amontoison closing after your explanation on allocations (linear solver, going to Mumps with MadNLP accomodates that, + some hidden alloc in native C++ code in the case of Ipopt)

[amontoison]

@jbcaillau The next major release of MadNLP.jl will use MUMPS by default (on CPU) as linear solver.