[klausb]

I want to use solveScalarField psi and const solveScalarField source from the two previous timesteps and use them in the current timestep in PCG.C for computations. This means I have to create four solveScalarFields in a way that they are accessible from within PCG.C . I want to use the functionality when running simpleFoam.

Where can/should I add the four solveScalarFields to store psi and source from the previous timesteps?

[dlahaye]

My extremely limited input:

1/ how exactly do you wish to modify the CG algorithm? Does PETSc offer more flexibility to implement what you are after, i.e., implement first in PETSc and subsequently employ PETSc4Foam?

2/ how do scalarField and solveScalarField differ?

3/ CG stores a vector for conjugate search directions (typically denoted by \vec{p} as in e.g. https://en.wikipedia.org/wiki/Conjugate_gradient_method ). Can you borrow from GMRES type ideas and extend the set of work vectors?

[klausb]

Originally Posted by dlahaye

My extremely limited input:

1/ how exactly do you wish to modify the CG algorithm? Does PETSc offer more flexibility to implement what you are after, i.e., implement first in PETSc and subsequently employ PETSc4Foam?

I want to implement an approach to improve the initial guess used when solving the linear system. The default is to use the result of the previous timestep. The approach requires information from at least two previous timesteps (psi and source).

2/ how do scalarField and solveScalarField differ?

I want to work with psi and source i.e. the solution and rhs vector of the linear system. These vectors are defined as solveScalarField see.: https://www.openfoam.com/documentati...8C_source.html

3/ CG stores a vector for conjugate search directions (typically denoted by \vec{p} as in e.g. https://en.wikipedia.org/wiki/Conjugate_gradient_method ). Can you borrow from GMRES type ideas and extend the set of work vectors?

It's not about reusing data within a loop e.g. related to restarting GMRES

It's really about finding a storage place to transfer, store and update information from multiple timesteps to be used and updated within PCG.C.

[dlahaye]

Not sure.

Make a child class of the currently in place symmetric matrix class, foresee two additional vector (aside from source and psi) and specialise PCG as a member function of this new class?

Note that CG converges non-linearly and that the effect of an improved initial guess might be limited.

Note that PETSc does provide tools for a deflated CG implementation. Deflation does more than merely improving the initial guess.

[klausb]

There's only #include "lduMatrix.H"

What I need are "global" variables in the e.g. simpleFoam solver to store psi and source vectors from t-1 and t-2 timesteps for processing and update in the current timestep from within the matrix solvers e.g. PCG, GAMG...