Interface for CFD solver

[shakedregev]

Summary

Ensure that Re::Solve can easily be called from a CFD application.

Rationale

Re::Solve has mostly been tested on linear systems extracted from solvers. We want to ensure it can be useful for production code.

Description

I suggest the following steps:

1. Build Re::Solve and run the tests and examples on CPU and both GPUs.
2. Extract systems from your CFD application in the format of the examples and benchmark and profile the different solvers for accuracy and run time. - This will help understanding the code and ensuring that the integration is useful.
3. Integrate Re::Solve within your solver. Add additional interfaces to Re::Solve if it is necessary/useful. Test and profile the different solvers.

Additional information

@pelesh - feel free to add anything.

[pelesh]

I think #400 is a good stepping stone for this activity.

[KakeruUeda]

Hi @pelesh, I think it would be great if SystemSolver exposed a public API to reset the preconditioner for iterative solvers, since the LHS matrix is updated every time step in CFD applications.

For example, SystemSolver could provide resetPreconditioner() that internally calls the reset functions (e.g., LinSolverDirectCuSparseILU0::reset(matrix::Sparse* A)).

I’d appreciate your thoughts on this.

[pelesh]

Hi @pelesh, I think it would be great if SystemSolver exposed a public API to reset the preconditioner for iterative solvers, since the LHS matrix is updated every time step in CFD applications.

For example, SystemSolver could provide resetPreconditioner() that internally calls the reset functions (e.g., LinSolverDirectCuSparseILU0::reset(matrix::Sparse* A)).

I’d appreciate your thoughts on this.

@KakeruUeda I agreee! We need to implement a generic preconditioner interface. Currently, we have a LU solver as a hard-wired interface with little flexibility. See #114.