Re: [Libmesh-devel] PetscVector::operator=

[Roy S. <roy...@ic...>]

On Thu, 24 Jun 2010, Jed Brown wrote:

> On Thu, 24 Jun 2010 15:59:22 +0200 (CEST), Tim Kroeger <tim...@ce...> wrote:
>
>> In my application, I am doing quite a lot of assignments between
>> NumericVector instances (which always are PetscVector instances in my
>> case), and since I am lazily most of the time working in opt mode, I
>> didn't realize that I was sometimes assigning vectors of different
>> type, that is in particular
>>
>> v = w
>>
>> where v is a parallel vector and w is a ghosted vector.
>
> This is value-assignment, right?

Right.

> There is no behavioral difference between a "parallel vector" and a
> "ghosted vector" except that you can only VecGhostGetLocalForm and
> VecGhostUpdate the latter.  You can certainly VecCopy between them.

Thanks for the update.

> Whether your operator= should support these mismatched types is entirely
> up to you.

I suspect there are only two reasons why we didn't support mismatched
types:

when that assertion was written we only had serial and parallel
vectors, a more extreme mismatch

we didn't know for sure that you could VecCopy between mismatched
types.

These reasons no longer apply.

> I could imagine imposing an artificial restriction just to
> maintain consistent types so that you don't accidentally end up calling
> VecGhostGetLocalForm on an plain vector (with work to figure out why
> it's a plain vector and not VecGhost).

As long as we don't copy the type over with operator=, this shouldn't
be a concern for PetscVector.  I'd say that we should at least enable 
"parallelvec = ghostedvec", if everyone agrees.  "parallelvec =
serialvec" and "ghostedvec = serialvec" would also be natural to
enable.  I'd just as soon eventually support every operator= combo,
but the "morelocaldatavec = lesslocaldatavec" type operations would
require us to add some code in addition to the plain VecCopy before
they'd have the semantics I'd want.
---
Roy