dolfin team mailing list archive

["Garth N. Wells" <gnw20@xxxxxxxxx>]

On 22 November 2011 21:58, Anders Logg <logg@xxxxxxxxx> wrote:
> On Tue, Nov 22, 2011 at 09:53:47PM +0000, Garth N. Wells wrote:
>> On 22 November 2011 21:50, Anders Logg <logg@xxxxxxxxx> wrote:
>> > On Tue, Nov 22, 2011 at 09:33:25PM +0000, Garth N. Wells wrote:
>> >> On 22 November 2011 21:30, Anders Logg <logg@xxxxxxxxx> wrote:
>> >> > On Tue, Nov 22, 2011 at 10:16:54PM +0100, Marie E. Rognes wrote:
>> >> >> On 11/22/2011 09:55 PM, Anders Logg wrote:
>> >> >> >On Tue, Nov 22, 2011 at 08:45:30PM +0000, Garth N. Wells wrote:
>> >> >> >>
>> >> >> >>On 21 Nov 2011, at 21:53, "Marie E. Rognes"<meg@xxxxxxxxx>  wrote:
>> >> >> >>
>> >> >> >>>
>> >> >> >>>On 21. nov. 2011, at 21:52, Anders Logg<logg@xxxxxxxxx>  wrote:
>> >> >> >>>
>> >> >> >>>>On Mon, Nov 21, 2011 at 08:46:13PM +0000, Garth N. Wells wrote:
>> >> >> >>>>>On 21 November 2011 13:07, Anders Logg<logg@xxxxxxxxx>  wrote:
>> >> >> >>>>>>On Sun, Nov 20, 2011 at 11:55:43PM +0100, Anders Logg wrote:
>> >> >> >>>>>>>On Sun, Nov 20, 2011 at 11:49:42PM +0100, Marie E. Rognes wrote:
>> >> >> >>>>>>>>
>> >> >> >>>>>>>>On 20. nov. 2011, at 23:31, Anders Logg<logg@xxxxxxxxx>  wrote:
>> >> >> >>>>>>>>
>> >> >> >>>>>>>>>Is anyone using the Function constructor that takes a vector as input
>> >> >> >>>>>>>>>argument?
>> >> >> >>>>>>>>>
>> >> >> >>>>>>>>>Function u(V, x);
>> >> >> >>>>>>>>>
>> >> >> >>>>>>>>Yes.
>> >> >> >>>>>>>Does it work? In parallel?
>> >> >> >>>>>>>
>> >> >> >>>>>>>Does it not work to instead use
>> >> >> >>>>>>>
>> >> >> >>>>>>>  x = u.vector()
>> >> >> >>>>>>>
>> >> >> >>>>>>>?
>> >> >> >>>>>>>
>> >> >> >>>>>>>If you need it, we should keep it but add an error message that it
>> >> >> >>>>>>>doesn't work in parallel, unless it does...
>> >> >> >>>>>>Any more input on this? There are several options:
>> >> >> >>>>>>
>> >> >> >>>>>>1. Remove this constructor
>> >> >> >>>>>>
>> >> >> >>>>>>2. Throw an error when running in parallel
>> >> >> >>>>>>
>> >> >> >>>>>>3. Check that the input vector makes sense
>> >> >> >>>>>>
>> >> >> >>>>>>The last one is problematic since I don't see an easy way to perform
>> >> >> >>>>>>the check, other than calling get_local and having it fail.
>> >> >> >>>>>>
>> >> >> >>>>>I haven't heard any reason why it can't be removed. We may need to fix
>> >> >> >>>>>assignment (re earlier discussion on assign) to just copy values and
>> >> >> >>>>>not the whole object so that a user can get the vector and then assign
>> >> >> >>>>>values to it without messing up the ghosting.
>> >> >> >>>>Sounds good, but I want to wait for Marie to comment before I remove
>> >> >> >>>>it. She is using it.
>> >> >> >>>>
>> >> >> >>>>Marie? Does it work for you to use x = u.vector()?
>> >> >> >>>>
>> >> >> >>>Probably. However removing the constructor would be changing parts of the basic interface, which I think is a bad idea.
>> >> >> >>>
>> >> >> >>>Add a warning if you want to deprecate it later.
>> >> >> >>>
>> >> >> >>Isn't the time to make an interface change now rather than later?
>> >> >>
>> >> >> I would say that the time to make an interface change before
>> >> >> 1.0 has passed: I see more value in sticking to
>> >> >> to what we have claimed, than in fixing this single instance.
>> >> >>
>> >> >> >True, but last time we discussed this was 1 hour or so before the
>> >> >> >release of 1.0-rc1. Now we have a whole week to 1.0-rc2... :-)
>> >> >> >
>> >> >> >Marie, can you check again if that constructor is necessary?
>> >> >>
>> >> >> I'm typically using it for the same as the dolfin la/eigenvalue demo
>> >> >> is using it for.
>> >> >> Do you have a replacement syntax available?
>> >
>> > Marie, I think this should work:
>> >
>> >  u = Function(V)
>> >  u.vector()[:] = x
>> >
>> > where x is the solution you get from the eigenvalue problem.
>> >
>> > Can you see if that works?
>> >
>>
>> It won't,  because of a bad flaw in the vector assignment. It will
>> make u.vector()[:] a copy of x (which has the wrong parallel layout) ,
>> when what we want is to assign just the values.
>
> I thought you argued before that was the correct behavior? (When we
> discussed the subfunction assignment last week.)
>

I argued that it is being done consistently, but not that it's right.
We need to distinguish between copying and object and assigning just
values of a vector (but not touching the layout, ghost values, etc) to
another vector of the same length.

> Is it much work to fix that before 1.0-rc2?
>

Proper assignment is too much work. I'm adding a work-around fix now
so the eigenvalue demo should be correct in parallel.

Garth

> --
> Anders
>
>
>> Garth
>>
>> >
>> >
>> >
>> >> >> That said, I'm not going to lose any sleep over this.
>> >> >
>> >> > Is everyone ok with throwing an error that it doesn't work in
>> >> > parallel?
>> >> >
>> >>
>> >> I don't think that is ideal.
>> >>
>> >> I building now with the constructors commented out to see how many
>> >> changes would be required.
>> >
>> >
>