Add DILU preconditioner

[jakobtorben]

This PR adds the Diagonal-ILU (DILU) preconditioner to OPM-Simulators.

DILU is a simplified version of the ILU(0) where only the diagonal terms are factorised, which means that we only need to store the diagonol for applying the preconditioner. More details on the method can be found here: https://netlib.org/linalg/html_templates/node64.html#SECTION00842200000000000000

[atgeirr]

This looks good, and should be merged eventually. I have some questions below, but mostly stylistic nitpicks. I have not really delved deeply into the numerics and checked them yet, though.

Congratulations on making your first OPM PR!

[atgeirr]

I do not understand why you need the Impl::asVector() and Impl::asMatrix()? I may be missing something, is this to handle the case of using scalars as block type?

[jakobtorben]

This was also inspired by code from Dune. Yes it wraps a scalar as a matrix and vector to support both:

  /** \brief Wrap a scalar as a 1-1-matrix */
  template<class T,
    std::enable_if_t<IsNumber<T>::value, int> = 0>
  auto asMatrix(T& t)
  {
    return ScalarMatrixView<T>{&t};
  }

I am not sure if we need this functionality for our case. I removed it for now and everything still works. I don't see it being used anywhere else in OPM-Simulators. I can add it back in if you want.

[blattms]

This is a cool addition.

What I am wondering is whether this is a good candidate for our new upstream strategy. I don't think there would be any issues in getting this into dune-istl.
I will soon create an opm-branch in DUNE modules where we need it. There we can merge patches before they are officially accepted. That way we could use it OPM right away if the user uses these branches.
We sometimes prepare short publications about releases (at least for the upcoming 2.9.1 release). If authors have contributed code and help with the publications they will become authors and get citations-

[jakobtorben]

@blattms That is cool to hear.
We have gotten promising results with this preconditioner so far. Using DILU as a preconditioner for BiCgStab it as fast as ILU0 and using DILU as a smoother in CPR-AMG it is even faster than using ILU0 as a smoother. We are looking into adding a parallel multicolour-DILU implementation as well, which we expect will make it even faster.

I can add this to dune-istl as well and contribute to a release publications if needed. Just let me know when the branch is set up and maybe give me a few pointers on how to integrate it. It is based on other preconditioners in DUNE, so I expect it will be quite easy to add.

[blattms]

You don't need to wait until the branch is set up. These are two indepent things. Your contribution should be a merge request to the master branch. In addition I will merge it to the opm branch (Of course I still have to set that up. (How can one be so slow as me 😅 )

Due to my slowness (still need to remove a clang problem in DUNE) 2.9.1 is still not out. So there might a very little chance to get that in even there. Depends on my charms and the release manager (that is waiting for my fix).

[atgeirr]

I may have some more comments, but I'll make a break now and post those I have so far.

[atgeirr]

jenkins build this please