[petsc-dev] error on preallocation business

[Barry Smith]

On Jan 21, 2012, at 2:53 PM, Jed Brown wrote:

> On Sat, Jan 21, 2012 at 14:46, Barry Smith <bsmith at mcs.anl.gov> wrote:
>    That patch is hideous, introduced more complexity.
> 
>    Why not  in MatSetUpPreallocation() add the one line
> 
>  if (!B->preallocated && B->ops->setuppreallocation) {
>    ierr = PetscInfo(B,"Warning not preallocating matrix storage\n");CHKERRQ(ierr);
>    ierr = (*B->ops->setuppreallocation)(B);CHKERRQ(ierr);
>    ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr);
> 
> This was my first inclination, but
> 
> 
> 1. DisAssemble_*() functions call preallocation routines again, but the values they use still aren't correct, so we have to avoid setting the options.

    The disassemble routines (in fact any PETSc library routines that call the preallocation routines) should ONLY call them if they are using correct values of preallocation otherwise they should not call them. So we should fix or remove all those bad uses of preallocation (preferably fix).  If you have the list of the ones that call with incorrect preallocation let us know.

> 
> 2. Users call preallocation directly (often through MatCreateMPIAIJ(), etc) with PETSC_DECIDE. In my opinion, they should see an error for not preallocating correctly. (But if you think they should have to say twice that they aren't preallocating correctly---by calling MatSetOption(), I can remove that uglyness.)

   I don't understand. With my model if they call MatCreateMPIAIJ() or MatMPIAIJSetPreallocation() it WILL set that flag and hence WILL bitch if they have used wrong values (even if they use PETSC_DECIDE. So it will do the right thing with my little fix; tell me why it won't.  Note if any of the preallocation routines are called it will not reset the flag in MatSetUpPreallocation().

> 
> 3. I didn't want to overwrite a user setting the option manually (perhaps with intent to preallocate).

        I already acknowledged this in my previous email. To me the tradeoff is monsterous code that Matt would hate vs this 

    In conclusion I think only 3 has any validity and that not enough for butt-ugly code.

    Barry

> 
> 
> that turns of the alloc error, note this only happens if the user did not preallocate.
> 
>   What do you say? Revert  that ugly code and add only one line of code?
> 
>   Barry
> 
> Note that it makes no sense for the user to not preallocate but then to set the error on alloc flag.
>