[petsc-dev] TS Terminology

[Jed Brown]

Matthew Knepley <knepley at gmail.com> writes:

> On Fri, Oct 20, 2017 at 11:58 AM, Emil Constantinescu <emconsta at mcs.anl.gov>
> wrote:
>
>> On 10/20/17 10:43 AM, Matthew Knepley wrote:
>>
>>> On Fri, Oct 20, 2017 at 11:22 AM, Emil Constantinescu <
>>> emconsta at mcs.anl.gov <mailto:emconsta at mcs.anl.gov>> wrote:
>>>
>>>     On 10/20/17 9:11 AM, Matthew Knepley wrote:
>>>
>>>         On Fri, Oct 20, 2017 at 9:23 AM, Emil Constantinescu
>>>         <emconsta at mcs.anl.gov <mailto:emconsta at mcs.anl.gov>
>>>         <mailto:emconsta at mcs.anl.gov <mailto:emconsta at mcs.anl.gov>>>
>>> wrote:
>>>
>>>              On 10/20/17 7:57 AM, Matthew Knepley wrote:
>>>
>>>                  I am confused by some of the terminology in TS. At the
>>> top
>>>                  level, IFunction appears to mean the entire equation
>>>
>>>                      F(u, u_t, x) = 0
>>>
>>>
>>>              Matt, page 141 of the manual: F(t, u, u_t) = G(t, u), and
>>>         not zero
>>>              on the RHS side. To make the interface general we allow
>>>         internally
>>>              for F:= F(t, u, u_t) - G(t, u) and then F=0.
>>>
>>>
>>>         This is not "internal". Its the toplevel interface:
>>>
>>>         https://bitbucket.org/petsc/petsc/src/63ae3ecac3af8ce782273a
>>> 76ad4152cddc2fd80a/src/ts/interface/ts.c?at=master&
>>> fileviewer=file-view-default#ts.c-884
>>>         <https://bitbucket.org/petsc/petsc/src/63ae3ecac3af8ce782273
>>> a76ad4152cddc2fd80a/src/ts/interface/ts.c?at=master&
>>> fileviewer=file-view-default#ts.c-884>
>>>
>>>         It computes F - G.
>>>
>>>
>>>     That's what it should do in some cases. The user provides either
>>>     ifunction or rhs funtion or both. The api to the solvers can take
>>>     care of this stuff automatically - that's what I meant by internal.
>>>     Different TS solvers can take different definitions of the funtions;
>>>     e.g., imex need both, beuler can take ifuntion and/or rhs function
>>>     but instead of writing beuler for both we choose the most general
>>>     case (ifunction) and compose the functions accordingly. The F - G is
>>>     transparent to the user. But somewhere the sausage needs to be made
>>>     and I think that is the right level because that is least likely to
>>>     change and least maintenance.
>>>
>>>
>>> I know what it does. I looked at the code. You are missing the point here.
>>>
>>> We cannot use the same word, IFunction, for two different things, F and
>>> F-G. The argument that is is not user facing is complete bullshit.
>>> The user inputs the points for ifunction, and can also call the toplevel
>>> interface.
>>>
>>
>> Matt, we do not. IFunction is F(t,u_t,u), RHS function is G(t,u). What we
>> solve is F=G and not F=0. Do you doubt that?
>>
>> When the user specifies IFunction it is that F; when the user specifies
>> RHS it is that G.
>>
>
> Nope. We use the word IFunction, specifically in the ifunction function
> pointer to mean
>
>   F
>
> but we use the word IFunction, specifically in TSComputeIFunction, to mean
>
>   F - G
>
> And, no its visible to everyone, not just "TS developers".

The "imex" flag is part of the TSComputeIFunction interface.  If that
flag is TRUE, then IFunction is exactly what the user provides.  If
FALSE, then the RHSFunction is subtracted off.  Should that be better
documented?

We could make a new interface TSComputeIFunctionMaybeMinusRHSFunction()
but I don't think it's necessary.  Without any such interface, the TS
implementations would each need to reproduce a bunch of vector and
matrix wrangling.

Note that TSComputeIFunction is very much like SNESComputeFunction,
which includes

  if (snes->vec_rhs) {
    ierr = VecAXPY(y,-1.0,snes->vec_rhs);CHKERRQ(ierr);
  }

Why haven't you complained about that?