dolfin team mailing list archive

["Benjamin Kehlet" <benjamik@xxxxxxxxxxxxxxxxxxxxx>]

> On Thu, Jul 16, 2009 at 08:26:21AM +0200, Benjamin Kehlet wrote:
>> >> On Thu, Jul 09, 2009 at 08:20:54AM +0100, Garth N. Wells wrote:
>> >>>
>> >>>
>> >>> Anders Logg wrote:
>> >>> > On Thu, Jul 09, 2009 at 08:05:42AM +0100, Garth N. Wells wrote:
>> >>> >>
>> >>> >> Anders Logg wrote:
>> >>> >>> On Thu, Jul 09, 2009 at 12:14:14AM +0200, DOLFIN wrote:
>> >>> >>>> One or more new changesets pushed to the primary dolfin
>> >>> repository.
>> >>> >>>> A short summary of the last three changesets is included below.
>> >>> >>>>
>> >>> >>>> changeset:   6458:8dacb335c9825380a2e614f5de2ae89f79131995
>> >>> >>>> tag:         tip
>> >>> >>>> user:        "Garth N. Wells <gnw20@xxxxxxxxx>"
>> >>> >>>> date:        Wed Jul 08 23:13:53 2009 +0100
>> >>> >>>> files:       demo/ode/reaction/cpp/main.cpp
>> >>> >>>> description:
>> >>> >>>> Comment out reaction ODE demo. Gives error when run.
>> >>> >>> What happens when you run it? Same as for the buildbot?
>> >>> >>>
>> >>> >>>   terminate called after throwing an instance of
>> >>> 'std::runtime_error'
>> >>> >>>   what():  *** Assertion (e1 != -1) [at
>> >>> >>>   dolfin/ode/MultiAdaptiveTimeSlab.cpp:926 in cg_feval()]
>> >>> >>>
>> >>> >>> This does not happen for me. It works fine and it's not because
>> I
>> >>> >>> don't trigger assertions, I checked.
>> >>> >>>
>> >>> >>> Any special configuration you are using? I'm running without
>> GMP.
>> >>> >>>
>> >>> >> It fails with GMP and works with GMP.
>> >>> >
>> >>>
>> >>> I was a bit hasty - fails with GMP and works without GMP.
>> >>>
>> >>> Garth
>> >>
>> >> ok, most likely some floating-point comparison that fails.
>> >>
>> >> I'll ask Benjamin Kehlet to look at it or fix it myself later
>> >> otherwise.
>> >
>> The problem is that in the reaction demo a number of real variables are
>> declared in main before the ode object is created.The precision of these
>> variables are incorrect since GMP hasn't been initialized at that time.
>> We need to initialize precision in GMP by calling mpf_set_default_prec()
>> before the first real is declared. We now do this in the constructor of
>> class GMPObject which ODE inherits.
>> The only way (I can think of) ensure absolutely, is to wrap class real
>> and
>> check in the constructor, but I guess the overhead by checking this
>> every
>> time a real is declared is unacceptable.
>> Another suggestion is to do the initalization in a function
>> dolfin::set_precision(uint prec) and force the user to call this
>> function.
>> We can check that this has been done when an ode object is created and
>> otherwise give an error message. The documentation and error message
>> must
>> state clearly  that this must be done _before_ creating any real
>> variables. The bad thing with such a solution is that the same user code
>> will not run both with and without GMP enabled, since we force the user
>> to
>> call set_precision, which doesn't make sense when running with the
>> normal
>> (double) precision.
>> Anders has suggested making this a parameter "floating-point precision"
>> in
>> Dolfin, but I can't see what should then trigger the call to
>> mpf_set_default_prec() before any reals are declared.
>
> Having a function named set_precision() sounds like the best solution.
>
> We can let this function return a warning instead of an error if GMP
> is not available. Then the demo would run both with and without GMP.
> I imaging that set_precision() will only be used in one or two ODE
> demos so it should be ok.
>
> Would that work?
>
That will work. I don't think the demos should give warnings when run
without GMP, but maybe we should write a demo demonstrating this
functionality (ie. solving some equation with a known exact solution to
see that the precision actually exceeds hardware precision).

Anders: I can have this ready in no time, but maybe we should wait to
after you're back from vacation. Or is there a release coming up?

Benjamin