Is DM and SX matrix creation thread safe?

[HaoruXue]

Is DM and SX matrix creation thread safe?

[jaeandersson]

Iirc, Function::call(const std::vector<DM> &arg, std::vector<DM> &res, ...) will not necessarily avoid the non-thread issues, for example if sparsity has to be projected. I think the easiest overload to use for thread safe function evaluation is arguments and results as std::map<std::string, std::vector< double >>

[fabmene]

This issue is already some months old, so you probably moved ahead with this already. But after dealing with a very similar issue, here's what I can tell:

• Creating DM, SX or MX is not thread safe
• Solvers may not be thread safe (IPOPT is, but even then the linear solver used by IPOPT might not be)
• Casadi function evaluations may create DMs behind the scenes if not set up exactly right.

The last point is the most tricky in my opinion, but this thread hints at it. The key takeaways:

Edit: Not the best approach. See comment below by @jaeandersson

• Don't use Function::operator(), but instead use Function::call(const std::vector<DM> &arg, std::vector<DM> &res, ...)
• This way you can pre-allocate all DM for arg and res
  • arg must contain in this order: x0, p, lbx, ubx, lbg, ubg, lam_x0, lam_g0
  • res must contain in this order: x, f, g, lam_x, lam_g, lam_p and will contain the result. Fill with anything.
  • They must all be of the correct size for your problem. If they are not, they are getting resized, which is not threadsafe

This is just from my trial and error and I don't have a deep understanding of how casadi works under the hood, so take with a grain of salt.

[fabmene]

Interesting, I didn't know you could avoid passing DMs altogether!