New integrator overhaul

[jaeandersson]

This is a new iteration (after #397) on the way AD is performed on integrator objects. This is necessary to enable support for multiple-shooting and (fast) high-order sensitivity analysis.

Prerequisites:

• Drop support for AD by OO, Drop support for operator overloading derivative calculation #884

Subtasks:

• Generalize generation of augmented DAE, Matrix-valued states in integrator not working #935 Problems differentiating MXFunction DAE #936
• More efficient sparsity propagation through integrator objects, Sparsity pattern propagation through integrators #934
• Access to algebraic variables, Change Integrator IO signature (Was: Obtaining the algebraic variable from an integrator) #504
• Solution statistics, Solution statistics, error flags from Sundials integrators #344
• Solution interpolation for fixed-step integrators, Solution interpolation with collocation integrator #245 Problem with CollocationIntegrator in Simulator context #539
• Exploit sensitivity analysis structure in SUNDIALS integrators, Exploit sensitivity analysis structure in SUNDIALS integrators #940
• Exploit Jacobian structure in augmented DAEs, Exploit Jacobian structure in augmented DAEs #941
• Make sure that integrators work properly together with parallelization, Error in Parallelizer #531
• Preconditioning bug in CVodes, Preconditioning not working in CVodes #815
• IDAS quadrature performance issue, quad_err_con poor performance with Idas #553
• Quadrature error control in SUNDIALS on by default, Make Sundials option "quad_err_con" true by default #546
• Avoid calculating the Jacobian in SUNDIALS when not needed, Jacobian calculated when not needed in CVodes/Idas #541
• Errors in CVodes backward Jacobian, Errors in CVodes backward Jacobian #540
• Refactor explicit Runge-Kutta integrators, Refactor the Runge-Kutta integrators #511
• Refactor collocation integrators, Refactor CollocationIntegrator #948

Cf. also #929.

Possible follow-ups: #945, #946, #947.

NOTE: Strikethrough tickets have been moved to #974.

[jaeandersson]

Branch "issue_794" created for this issue. @jgillis Note that a major change of integrators is on the way. If you want to make changes to the integrators, better make it to this branch. Otherwise it will surely lead to conflicts.

[jgillis]

Bump "Retreiving algorithmic states when using simulator"

(Anne Mai)

[jaeandersson]

Does not make sense to address algebraic variable outputs before refactoring. Will look at it when I'm happy with the overall design. Will work on this issue after #884, which is related.

[ghorn]

If you need algebraic variables, but direct coll is too expensive, could you just solve the residual at one additional point per timestep to get the alg vars manually?

[jaeandersson]

Yes although initial guess could be an issue.

[jaeandersson]

I don't think anymore it makes sense to implement this in a separate branch. The changes will be evolutionary and not revolutionary with few or no syntax changes. Therefore, I'll close this issue and deal deal with the specific changes in dedicated issues.

[jaeandersson]

I'll reopen the issue, since it's good to have all the changes summarized in one place.

[jaeandersson]

The refactoring of the generation of the augmented DAE is now complete. While the new implementation is not necessarily better for SXFunction, it should scale better and work with callback functions of any type. Added an option "expand_augmented" which can be set to "false" to achieve a possibly smaller memory footprint. If the numerical efficiency of MXFunction improves further, this option can be removed.

[jaeandersson]

Closing this issue, postponing remaining tasks to #974.

[jaeandersson]

Branches issue_794 and issue_794b removed.