Previous  Up  Next

Module Ida.Spils

module Spils: sig .. end

Scaled Preconditioned Iterative Linear Solvers.

include Sundials_LinearSolver.Iterative

Preconditioners

type 'd prec_solve_fn = (unit, 'd) Ida.jacobian_arg -> 'd -> 'd -> float -> unit

Callback functions that solve a linear system involving a preconditioner matrix. In the call prec_solve_fn jac r z delta, jac is a Ida.jacobian_arg with no work vectors, r is the right-hand side vector, z is computed to solve $Pz = r$, and delta is the input tolerance. $P$ is a preconditioner matrix, which approximates, however crudely, the Jacobian matrix $\frac{\partial F}{\partial y} + \mathtt{arg.jac\_coef}\frac{\partial F}{\partial\dot{y}}$. If the solution is found via an iterative method, it must satisfy $\sqrt{\sum_i (\mathit{Res}_i \cdot \mathit{ewt}_i)^2} < \mathtt{delta}$, where $\mathit{Res} = r - Pz$ and $\mathit{ewt}$ comes from Ida.get_err_weights.

Raising Sundials.RecoverableFailure indicates a recoverable error. Any other exception is treated as an unrecoverable error.

The elements of jac, r, and z should not be accessed after the function has returned.

type 'd prec_setup_fn = (unit, 'd) Ida.jacobian_arg -> unit

Callback functions that preprocess or evaluate Jacobian-related data need by Ida.Spils.prec_solve_fn. The sole argument is a Ida.jacobian_arg with no work vectors.

Raising Sundials.RecoverableFailure indicates a recoverable error. Any other exception is treated as an unrecoverable error.

The elements of the argument should not be accessed after the function has returned.

type ('d, 'k) preconditioner = ('d, 'k) SpilsTypes.preconditioner

Specifies a preconditioner and its callback functions. The following functions and those in Ida_bbd construct preconditioners.

The Ida.Spils.prec_solve_fn is mandatory. The Ida.Spils.prec_setup_fn can be omitted if not needed.

val prec_none : ('d, 'k) preconditioner

No preconditioning.

val prec_left : ?setup:'d prec_setup_fn ->
       'd prec_solve_fn -> ('d, 'k) preconditioner

Left preconditioning. $Pz = r$, where $P$ approximates, perhaps crudely, $J = \frac{\partial F}{\partial y} + c_j\frac{\partial F}{\partial\dot{y}}$.

Solvers

type 'd jac_times_setup_fn = (unit, 'd) Ida.jacobian_arg -> unit

Callback functions that preprocess or evaluate Jacobian-related data needed by the jac_times_vec_fn. In the call jac_times_setup_fn arg, arg is a Ida.jacobian_arg with no work vectors.

Raising Sundials.RecoverableFailure indicates a recoverable error. Any other exception is treated as an unrecoverable error.

The elements of arg should not be accessed after the function has returned.

type 'd jac_times_vec_fn = ('d Ida.double, 'd) Ida.jacobian_arg -> 'd -> 'd -> unit

Callback functions that compute the Jacobian times a vector. In the call jac_times_vec_fn arg v jv, arg is a Ida.jacobian_arg with two work vectors, v is the vector multiplying the Jacobian, and jv is the vector in which to store the result—$\mathtt{jv} = J\mathtt{v}$.

Raising Sundials.RecoverableFailure indicates a recoverable error. Any other exception is treated as an unrecoverable error.

Neither the elements of arg nor v or jv should be accessed after the function has returned.

val solver : ('m, 'd, 'k, [> `Iter ]) Sundials.LinearSolver.t ->
       ?jac_times_vec:'d jac_times_setup_fn option *
                      'd jac_times_vec_fn ->
       ?jac_times_res:'d Ida.resfn ->
       ('d, 'k) preconditioner -> ('d, 'k) Ida.linear_solver

Create an Ida-specific linear solver from a generic iterative linear solver.

The jac_times_res argument specifies an alternative DAE residual function for use in the internal Jacobian-vector product difference quotient approximation. It is incorrect to specify both this argument and jac_times_vec.

NB: a jac_times_setup_fn is not supported in Config.sundials_version < 3.0.0.

NB: a jac_times_res function is not supported in Config.sundials_version < 5.3.0.

Solver parameters

val set_eps_lin : ('d, 'k) Ida.session -> float -> unit

Sets the factor by which the Krylov linear solver's convergence test constant is reduced from the Newton iteration test constant. This factor must be >= 0; passing 0 specifies the default (0.05).

val set_ls_norm_factor : ('d, 'k) Ida.session -> float -> unit

Sets the factor for converting from the integrator tolerance (WRMS norm) to the linear solver tolerance (L2 norm). That is, $\mathit{tol}_{\mathsf{L2}} = \mathit{fact}\cdot\mathit{tol}_{\mathsf{WRMS}}$ . The given value is used directly if it is greater than zero. If it is zero (the default), then the square root of the state vector length is used. If it is less than zero, then the square root of the dot product of a state vector full of ones with itself is used.

val set_linear_solution_scaling : ('d, 'k) Ida.session -> bool -> unit

Enables or disables scaling of the linear system solution to account for a change in $\gamma$ in the linear system. Linear solution scaling is enabled by default when a matrix-based linear solver is attached.

val set_increment_factor : ('d, 'k) Ida.session -> float -> unit

Sets the increment factor (dqincfac) to use in the difference-quotient approximation.

Specifically, the product $Jv$ is approximated by $Jv = \frac{1}{\sigma}\left( F(t, \tilde{y}, \tilde{y}') - F(t, y, y') \right)$ . where $\tilde{y} = y + \sigma v$ , $\tilde{y}' = y' + c_j \sigma v$ , $c_j$ is a BDF parameter proportional to the step size, $\sigma = \sqrt{N} \mathtt{dqincfac}$ , and $N$ is the number of equations in the DAE system.

Solver statistics

val get_work_space : ('d, 'k) Ida.session -> int * int

Returns the sizes of the real and integer workspaces used by the spils linear solver.

val get_num_lin_iters : ('d, 'k) Ida.session -> int

Returns the cumulative number of linear iterations.

val get_num_lin_conv_fails : ('d, 'k) Ida.session -> int

Returns the cumulative number of linear convergence failures.

val get_num_prec_evals : ('d, 'k) Ida.session -> int

Returns the number of calls to the setup function.

val get_num_prec_solves : ('d, 'k) Ida.session -> int

Returns the cumulative number of calls to the preconditioner solve function.

val get_num_jtsetup_evals : ('d, 'k) Ida.session -> int

Returns the cumulative number of calls to the Jacobian-vector setup function.

val get_num_jtimes_evals : ('d, 'k) Ida.session -> int

Returns the cumulative number of calls to the Jacobian-vector function.

val get_num_lin_res_evals : ('d, 'k) Ida.session -> int

Returns the number of calls to the residual callback for finite difference Jacobian-vector product approximation. This counter is only updated if the default difference quotient function is used.

Low-level solver manipulation

The Ida.init and Ida.reinit functions are the preferred way to set or change preconditioner functions. These low-level functions are provided for experts who want to avoid resetting internal counters and other associated side-effects.

val set_preconditioner : ('d, 'k) Ida.session ->
       ?setup:'d prec_setup_fn -> 'd prec_solve_fn -> unit

Change the preconditioner functions.

val set_jac_times : ('d, 'k) Ida.session ->
       ?jac_times_setup:'d jac_times_setup_fn ->
       'd jac_times_vec_fn -> unit

Change the Jacobian-times-vector function.

NB: the jac_times_setup argument is not supported in Config.sundials_version < 3.0.0.

val clear_jac_times : ('d, 'k) Ida.session -> unit

Remove a Jacobian-times-vector function and use the default implementation.