[petsc-users] local/global Vec for SNES function/jacobian

[Smith, Barry F.]

> On Jan 9, 2018, at 12:53 AM, Matteo Semplice <matteo.semplice at unito.it> wrote:
> 
> Dear all,
> 
>     I could not find info on the manual but from the examples and my own trying I guess that the vectors involved in the SNESfunction should be global ones and that, if one needs ghosts values to compute the function and/or the jacobian, should do a GlobalToLocal communication inside the function. Correct?

  Absolutely

> (I get errors of inconsistent dimensions between Vecs and Mats if I call SNESSolve(snes,b, localU)...)
> 
>     Since I need the ghost values both in the function and in the jacobian evaluation, would it be safe (and worthwhile) to do define a local vector in the ctx, perform the communication in the function and using the vec in the ctx in the jacobian evaluation (without performing again the communication in the SNESJacobianFunction?)... Put in other words, is it guaranteed that each call to the jacobian function is preceded by a call to the function with the same x argument?

  I would not rely on this. It is an unnecessary and risky optimization. Just treat the function and Jacobian evaluations as independent operations.

  Barry

> 
> Thanks,
>     Matteo Semplice