Dear Carlo,
thank you for your suggestions.
Carlo De Falco wrote:
>On 16 Jun 2016, at 17:25, Francesco Faccio <address@hidden> wrote:
>
>> Sundials uses its own types: realtype and N_Vector. An N_Vector is a vector of realtype, while a realtype can be both a float, a double or a long double, depending on how Sundials has been built
(default type is double).
>
>I think you should make sure at compile time that realtime in sundials is double and the
>use double instead of realtime in the code
I think I can write a macro inside m4/acinclude.m4 and call it from configure.ac to check this.
>I'm not sure Matlab supports solving DAEs/IDEs using floats for state vector and jacobian,
>can you check?
Matlab allows the user to use values of single or double precision.
>According to the documentation, IDA can apprroximate the jacobian itself if needed:
IDA can approximate only a modified Jacobian in the form J = DF/DY + cj*DF/DYP, where cj is a scalar proportional to the inverse of the step size (cj is computed by IDA's solver).
In the Dense case, according to Matlab code, there can be 5 different situations and this is how we could solve them:
- the user doesn't supply a Jacobian: IDA approximates it by difference-quotients
- the user provides a modified Jacobian { [ DF/DY ], [ DF/DYP ] } : I can pass it to IDA
- the user provides { [ DF/DY ], [ ] } : I need to approximate DF/DYP before passing it to IDA
- the user provides { [ ], [ DF/DYP] } : I need to approximate DF/DY before passing it to IDA
- the user provides a function returning [ DF/DY, DF/DYP ] : I can pass it to IDA
Since IDA makes all its approximations inside IDASOLVE I will need to interface with an other library (or to modify Sundials' code).
Francesco