New ParU and PARDISO perfect foresight solvers authored by Sébastien Villemot's avatar Sébastien Villemot
......@@ -25,6 +25,10 @@ This page documents the changes between Dynare versions (including the unstable
+ The old default, using an incomplete LU decomposition, remains available under `precondition=incomplete_lu`
+ The `first_iter_lu` and `block_diagonal_lu` preconditioners deliver a significant performance improvement over the old default (`incomplete_lu`) and make this algorithm competitive and in some cases faster than other perfect foresight solvers
+ Other new options that control the behaviour of the iterative solvers: `iter_tol`, `iter_maxit`, `gmres_restart`, `block_diagonal_lu_maxlu`, `block_diagonal_lu_nperiods`, `block_diagonal_lu_nlu`, `block_diagonal_lu_relu`
- New perfect foresight solvers (4cf13f545d0095ed7cfbf056b1100dedf8f1d1c0)
+ ParU from SuiteSparse, a direct sparse LU solver, available under the `stack_solve_algo=8` option
+ Panua PARDISO direct sparse LU solver, available under the `stack_solve_algo=9` option (license file has to be provided by the user)
+ Panua PARDISO direct-iterative solver using CGS, available under the `stack_solve_algo=10` option (license file has to be provided by the user)
- Significant performance improvement of higher-order perturbation solution under Windows (#1962)
- Collect and store optimization information: `dynare_minimize_objective` now returns runtime, iterations, function evaluations, exitflag, and messages from all optimizers (!2375)
- Option values `solve_algo={6,7,8}` and `stack_solve_algo={2,3}` are now available without `block` nor `bytecode` (e14a49211067b27114d8173b16d4cb8f7595b0c3, aa263e6a2adc9414142551a01bbcc46abb6a3894)
......
......