[petsc-dev] assuming sizeof(PetscInt) <= sizeof(PetscScalar)

[Barry Smith]

On Feb 8, 2014, at 6:12 PM, Peter Brune <brune at mcs.anl.gov> wrote:

> I'll get this stuff out of MatCoarsen and PCGAMG.  I will note that users/developers wouldn't do this if there were a good way of taking indices or proc/index pairs out of scatters.

  Then add an API to do that.

  Barry

> 
> One thought on —with-64-bit-indices —with-precision=single: reducing over 4 billion floats sounds like a bad time.
> 
> - Peter
> 
> 
> On Sat, Feb 8, 2014 at 5:52 PM, Matthew Knepley <knepley at gmail.com> wrote:
> On Sat, Feb 8, 2014 at 5:40 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
>    It appears people have been assuming sizeof(PetscInt) <= sizeof(PetscScalar) right and left in the code (to take advantage of VecScatter functionality). This is most definitely not true with —with-64-bit-indices —with-precision=single
> 
> MatGetBrowsOfAoCols_MPIAIJ
> PCGAMGProlongator_Classical_Standard
> maybe others
> 
> VecScatters are for moving Vec elements around, they are not a replacement for MPI calls!
> 
> Please don’t do this in the future.
> 
> Please please please use PetscSF if you have regular layouts, and PetscSection for irregular layouts.
> 
>   Matt
>  
> 
>    Barry
> 
> 
> 
> 
> -- 
> What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> -- Norbert Wiener
>