On 22 November 2011 21:50, Anders Logg <[email protected]> wrote: > On Tue, Nov 22, 2011 at 09:33:25PM +0000, Garth N. Wells wrote: >> On 22 November 2011 21:30, Anders Logg <[email protected]> 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"<[email protected]> wrote: >> >> >> >> >> >>> >> >> >>>On 21. nov. 2011, at 21:52, Anders Logg<[email protected]> wrote: >> >> >>> >> >> >>>>On Mon, Nov 21, 2011 at 08:46:13PM +0000, Garth N. Wells wrote: >> >> >>>>>On 21 November 2011 13:07, Anders Logg<[email protected]> 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<[email protected]> 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. Garth > -- > Anders > > > >> >> 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. > > _______________________________________________ Mailing list: https://launchpad.net/~dolfin Post to : [email protected] Unsubscribe : https://launchpad.net/~dolfin More help : https://help.launchpad.net/ListHelp
