Scalar dynamic evaluation

[GitPaean]

The main purpose of this PR is to be able to create a DynamicEvaluation directly/implicitly from a constant, like 0.0 or 1.0 or 0.5, without specifying the number of the derivatives.

For example, for the following code,

EvalWell cq_s_zfrac_effective{this->primary_variables_.numWellEq() + Indices::numEq, 0.0};

now we can do

EvalWell cq_s_zfrac_effective{0.0};

To do this, we represent constants with DynamicEvaluation without derivatives. And the operators are updated to incorporate this change.

The main motivation was that we have a lot of material related calculation using implicit conversion from scalar values. When we tried to do more material related calculation (similar to flash calculation for black oil) inside the wellbore in Standard Well, it appeared that lacking of implicit conversion caused issues. We can (tried to) update the material calculations to handle DyanmiceEvaluation, while it will affect many files and the approach introduced this PR is the more confined manner.

[GitPaean]

jenkins build this please

[GitPaean]

jenkins build this please

[Copilot]

Pull Request Overview

This PR adds support for scalar dynamic evaluation by implementing constant evaluation handling in the DynamicEvaluation class. The changes allow operations between evaluations with different derivative sizes, where one can be a constant (size 0).

• Adds a new constructor for creating constant evaluations with zero derivatives
• Implements logic to handle operations between evaluations of different sizes when one is constant
• Updates createConstant method to return a proper constant evaluation instead of throwing

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 12 comments.

File	Description
opm/material/densead/DynamicEvaluation.hpp	Adds constant evaluation support and mixed-size operation handling
opm/material/common/FastSmallVector.hpp	Adds expandSize method to support dynamic resizing of evaluation storage
jenkins/pre-build.sh	Comments out code generation script (temporary test change)

[GitPaean]

jenkins build this please

[GitPaean]

jenkins build this please

[GitPaean]

It looks like some rounding error or other mistakes jumping in caused regression. Working on recovering it.

[GitPaean]

jenkins build this please

[atgeirr]

Just commenting on the expandSize() method now.

I generally support the improved flexibility for the DynamicEvaluation, but have not looked at that yet.

[GitPaean]

jenkins build this please

[GitPaean]

jenkins build this please

[GitPaean]

jenkins build this please

[GitPaean]

With some discussion, it looks like this is an helpful/useful things to do. I am removing the [test] from the title of the PR.

@atgeirr , I have updated the PR, please have a look. I will migrate the implementation to the .py file.

[GitPaean]

I have updated the genEvalSpecializations.py to generate the changes introduced in this PR and marking the PR is ready for review.

I updated the description for this PR and mark it as ready for review.

[GitPaean]

jenkins build this please

[GitPaean]

I will make an accompanying PR in the opm-simulators to test how it works. Currently only test it with a development PR while not tested the usage in the master branch, so I am converting this PR to be draft again.

[GitPaean]

The failure (https://ci.opm-project.org/job/opm-simulators-PR-builder/8802/testReport/junit/(root)/mpi/compareECLFiles_flow_GASCONDENSATE_VAPWAT_PRECSALT_REGRESSION/) is due to the following code

in here is a DynamicEvalution initialized with scalar 0., so it does not have any derivatives. We will not be able to get any derivative to copy for the following code.

    template <class EvalWell>
    Eval restrictEval(const EvalWell& in) const
    {
        Eval out = 0.0;
        out.setValue(in.value());
        for (int eqIdx = 0; eqIdx < Indices::numEq; ++eqIdx) {
            out.setDerivative(eqIdx, in.derivative(eqIdx));
        }
        return out;
    }

[GitPaean]

jenkins build this opm-simulators=6520 please

[GitPaean]

jenkins build this opm-simulators=6520 please

[GitPaean]

jenkins build this opm-simulators=6520 please

[GitPaean]

jenkins build this opm-simulators=6520 please

[atgeirr]

The comment on the DynamicEvaluation is made in that source file because it was simpler for me, but you would have to make changes in the generating .py file instead.

[atgeirr]

I think that instead of extending this object's derivatives with the appendDerivativesToConstant() function, you could do:

const T value_copy = value();
*this = other;
*this /= value_copy;

I think this pattern is a lot simpler, and should be done for all the arithmetic operations here.

[GitPaean]

I think the original comment was wrong. if this is a constant, we divide all values and derivatives by the value.

I corrected the comment to be // if *this is a constant, extend to hold derivatives before division.

Sorry about that.

your suggestion will be the other way around,

other /= *this

Let me know if I read it wrong. I worked on it some time ago, I might be confused here.

[atgeirr]

No you are right, sorry for adding confusion! I do think that for the other operators the pattern I suggested would work, but of course not for division, a/b is not equal to b/a, I do not know what hit me.

[GitPaean]

jenkins build this opm-simulators=6520 please

[GitPaean]

Thanks for the approval.

The PR can go in independently. Let me test it only itself without the downstream PR.

[GitPaean]

jenkins build this please

[GitPaean]

jenkins build this please

[GitPaean]

jenkins build this please

[GitPaean]

the nonnull warming is silenced.

[GitPaean]

jenkins build this please

[GitPaean]

approved and jenkins passed. I am getting the PR in now.