[WIP] Harmonize solvers #1030
[WIP] Harmonize solvers #1030
Commits on Sep 8, 2023
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Remove logging hook for exceptions
Logging prevents app from reporting exception to commandline
-
Remove preconditioner if
precon=NonePETSc applies preconditioners by default
-
-
-
-
-
-
-
-
-
Change arguments and don't loop over solver suites
Different solver suites have different preconditioners, so too complicated to try to capture all in a single loop. Choose suite and preconditioner at calling level.
-
-
-
Swap order of looping over solvers and looping over sizes
Allow specifying minimum and maximum sizes
-
-
-
-
-
-
Change solver selection to cascade of ifs
Trying to be too clever with dict means that it doesn't work for all solver suites
-
-
-
-
-
-
-
-
-
-
Save results in
.npzinstead of.tsvhttps://numpy.org/doc/stable/reference/generated/numpy.savez.html Soooooo much faster (and [documented!](https://numpy.org/doc/stable/reference/generated/numpy.lib.format.html#module-numpy.lib.format))
-
Log much more information about FiPy environment
Remove obsolete vitals module (svn, really???)
-
Attempt to make problem less stiff
- Use harmonic, rather than arithmetic, mean - Enable setting other (non-zero) Dirichlet boundary conditions
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Shift diffusion problem definition to notebook
Generate diffusion.py from diffusion.ipynb
-
-
-
-
-
-
-
-
Commits on Sep 29, 2023
Commits on Oct 10, 2023
Commits on Oct 19, 2023
Commits on Oct 26, 2023
-
CSR requires *slow* conversion from IJV. CRV is "Undocumented.", but turns out to be what we want.
Commits on Oct 31, 2023
Commits on Nov 2, 2023
Commits on Nov 3, 2023
-
Make solver-independent Convergence and Divergence classes.
Solvers subclass these. Simplifies documentation.
-
-
-
Commits on Nov 7, 2023
-
-
Remove
max_iterationsfrom ConvergenceRedundant to information in Solver, which Convergence points to.
-
-
-
Commits on Nov 8, 2023
-
-
-
-
-
-
-
Change default solver tolerance to 1e-5
- SciPy and Trilinos can fail with tolerance=1e-10 (SCIPY_MAXIT or AZ_loss, respectively) because they are unable to make the residual any smaller than O(1e-9). - tolerance=1e-5 is the default for PETSc and SciPy. - pyamgx defaults to 1e-12. - PySparse does not specify, but has examples that illustrate 1e-12. - Trilinos does not specify, but has examples that illustrate 1e-8.
Commits on Nov 9, 2023
-
-
Add "default" preconditioner for PETSc
Use PETSc's own defaults https://petsc.org/main/manualpages/PC/PCCreate/#note
-
Commits on Nov 10, 2023
-
Consolidate Solver default parameters
Eliminate cumbersome telephone game for a small number of changed defaults in subclasses. Previous was error-prone and tedious to change global defaults. Contrary to [official guidance](https://docs.python-guide.org/writing/gotchas/#what-you-should-do-instead), use `"default"` instead of `None` to denote default argument. Semantics of using None are confusing when there are times we want no preconditioner or no tolerance limit and other times we want a default preconditioner or default tolerance limit.
Commits on Nov 11, 2023
Commits on Nov 18, 2023
Commits on Nov 21, 2023
Commits on Dec 4, 2023
Commits on Dec 14, 2023
Commits on Dec 22, 2023
Commits on Dec 23, 2023
Commits on Dec 24, 2023
Commits on Dec 27, 2023
Commits on Dec 29, 2023
Commits on Jan 2, 2024
-
Standardize that
Solver._Lxbreturns a FiPy _SparseMatrixPreviously inconsistent about whether `L` was a `_SparseMatrix` or an underlying suite sparse matrix. `x` and `b` continue to be appropriate vectors for the suite.
-
Commits on Jan 3, 2024
-
Refactor
_solve()and_solve_()`Solver._solve()` is common and has access to FiPy objects. `Solver._solve_()` is defined for each solver suite and takes L, x, b suitable for particular solver suite.
-
Commits on Jan 5, 2024
-
-
`Norm2()` returns a 1-element array for some dain-bramaged reason.
-
-
-
-
-
Make input4thOrder1D use LU for all solver suites
All (most) suites are flakey with this problem. (Trilinos only fails in parallel if the rest of the examples/diffusion tests have already been run.)
-
Commits on Jan 9, 2024
-
Prevent solver from getting "stuck"
RHS >> initial residual, so solver never does anything with (new) default criterion for convergence.
-
-
-
-
Still passes and doesn't throw obnoxious warnings from PETSc or Trilinos
Commits on Jan 10, 2024
Commits on Jan 11, 2024
-
-
Adjust tolerances to account for relaxed default solver tolerance
Default solver tolerances have been relaxed from 1e-10 to 1e-5 to agree with default of most solver suites. This should save on compute time and, in most cases, this is adequate for existing tests to pass. Where it is not, either adjust the solver or adjust the test.
Commits on Jan 12, 2024
Commits on Jan 19, 2024
-
Introduce Timer context manager
Originally tried to make it do logging, too, but I couldn't figure out how to get it to log in the correct scope. Separating timing from logging substantially simplifies the logic at a modest cost in LoC at point of use.
-
Commits on Apr 21, 2024
Commits on Apr 22, 2024
-
-
Turn constraint contributions into dictionary keyed on variable
Terms should, in principle, be solvable for different variables.
-
-