dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Nov 12, 2012 at 08:37:47AM +0000, Garth N. Wells wrote:
> On Mon, Nov 12, 2012 at 8:31 AM, Anders Logg <logg@xxxxxxxxx> wrote:
> > On Sun, Nov 11, 2012 at 09:33:01PM +0000, Garth N. Wells wrote:
> >> On Sun, Nov 11, 2012 at 7:36 PM, Anders Logg <logg@xxxxxxxxx> wrote:
> >> > On Sun, Nov 11, 2012 at 10:22:12AM +0000, Garth N. Wells wrote:
> >> >> On Fri, Oct 26, 2012 at 6:22 AM, Anders Logg <logg@xxxxxxxxx> wrote:
> >> >> > On Mon, Oct 22, 2012 at 10:32:11AM +0100, Garth N. Wells wrote:
> >> >> >> We have discussed briefly in the past changing from unsigned int
> >> >> >> (typedef uint) to std::size_t. Starting to solve some really big
> >> >> >> problems and some changes in Trilinos make it a good time to bring
> >> >> >> this up again. Any thoughts or objections to moving to std::size_t
> >> >> >> from uint?
> >> >> >
> >> >> > I think this would be a good idea.
> >> >> >
> >> >>
> >> >> I've started making some unsigned int -> std::size_t changes as I
> >> >> restructure mesh partitioning.
> >> >>
> >> >> > I suggest we keep the uint typedef and make it point to size_t.
> >> >> >
> >> >>
> >> >> I think we should use std::size_t and not uint. std::size_t is already
> >> >> a typedef and it conveys an intention: big enough for the largest
> >> >> array that can be allocated on a machine.  Also, it's not a question
> >> >> of unsigned int or std::size_t - there are places for both.
> >> >
> >> > So we will keep dolfin::uint for stuff like component indices and
> >> > other small integers, and use size_t for everything that can
> >> > potentially be large?
> >>
> >> Yes. I lean towards using 'unsigned int' instead of 'dolfin::uint'.
> >
> > Why? To minimize internal typedefs?
> >
>
> Yes. Typing 'unsigned int' in full doesn't bother me.

I don't feel strongly about it, as long as we're consistent.

> >> > How about the Mesh? Should we use size_t for stuff like mesh
> >> > connectivity?
> >> >
> >>
> >> If it can potentially be big, then it should be std::size_t.
> >
> > Is the assumption that global dof numbers need size_t while for local
> > entity indices (to a process) it's enough with uint?
> >
>
> I would suggest using std::size_t for local indices.
>
> I've used unsigned int for things like topological and geometric
> dimensions, number of connected entities, number of entities per cell,
> etc.

Is there a performance/memory hit?

--
Anders