dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Fri, Feb 11, 2011 at 11:14:33AM -0800, Johan Hake wrote:
> On Friday February 11 2011 11:06:15 Anders Logg wrote:
> > On Fri, Feb 11, 2011 at 10:10:51AM -0800, Johan Hake wrote:
> > > On Friday February 11 2011 09:54:39 Anders Logg wrote:
> > > > On Fri, Feb 11, 2011 at 08:08:32AM -0800, Johan Hake wrote:
> > > > > On Friday February 11 2011 01:16:31 Anders Logg wrote:
> > > > > > How is DOLFIN_EPS treated in Expressions and boundary conditions
> > > > > > (using compile_subdomains)?
> > > > > >
> > > > > > I thought that DOLFIN_EPS was known to the JIT compiler.
> > > > > >
> > > > > > Here are somethings I noted while debugging a code:
> > > > > >
> > > > > > 1. f = Expression("sin(x[0] + DOLFIN_EPS)") does not compile.
> > > > > >
> > > > > > The log file pointed to by the JIT compiler says "syntax error" and
> > > > > > the generated code in .i file found in the same directory contains
> > > > > > this:
> > > > > >
> > > > > > class Expression_5d66d4e2b02f54f6088fc61767e40073: public
> > > > > > Expression {
> > > > > >
> > > > > > public:
> > > > > >   double DOLFIN_EPS;
> > > > > >   Expression_5d66d4e2b02f54f6088fc61767e40073():Expression()
> > > > > >   {
> > > > > >
> > > > > >     DOLFIN_EPS = 0.0;
> > > > > >
> > > > > >   }
> > > > > >
> > > > > >   void eval(dolfin::Array<double>& values, const
> > > > > >   dolfin::Array<double>& x) const
> > > > > >   {
> > > > > >
> > > > > >     values[0] = sin(x[0] + DOLFIN_EPS);
> > > > > >
> > > > > >   }
> > > > > >
> > > > > > };
> > > > > >
> > > > > > So there are two problems here:
> > > > > >
> > > > > > 1.a) The JIT compiler fails to compile the Expression f which looks
> > > > > > like a valid expression to me
> > > > > >
> > > > > > 1.b) DOLFIN_EPS is set to zero.
> > > > >
> > > > > Now it doesn't.
> > > >
> > > > Thanks!
> > > >
> > > > > > 2. When trying something similar with compile_subdomains, it seems
> > > > > > that
> > > > > >
> > > > > >   compile_subdomains("x[0] < DOLFIN_EPS")
> > > > > >
> > > > > > compiles and produces correct code. In particular the .i file does
> > > > > > not contain DOLFIN_EPS = 0.0.
> > > > >
> > > > > subdomain treats DOLFIN_EPS and on_boundary seperately.
> > > > >
> > > > > > But the magic for creating variables is a bit too clever. I had a
> > > > > > typo which said "DOFLIN_EPS" instead of "DOLFIN_EPS" which
> > > > > > compiles without any warnings and DOFLIN_EPS is set to zero.
> > > > > > That's a litle dangerous.
> > > > >
> > > > > You cannot blame your typos on compile expressions ;)
> > > >
> > > > I just did! ;-)
> > > >
> > > > Would it be possible (and desirable) to require default values for all
> > > > variables in an Expression?
> > > >
> > > > f = Expression("a*x[0]", a=2.0)
> > > > f.a = 3.0
> > >
> > > A good point. It would probably require to hide the kwarg: element, cell,
> > > degree, into a **kwargs, making them less transparent. We could also add
> > > the default value **kwargs after the above mention kwargs. If we go for
> > > this we
> > >
> > > can end up with syntax like:
> > >   f = Expression("a*x[0]", element=some_element, a=2.0)
> > >
> > > instead of:
> > >   f = Expression("a*x[0]", a=2.0, element=some_element)
> > >
> > > I think the last syntax makes more sense, but it requires more hidden
> > > logics.
> >
> > If it's important to keep the default argumets visible, you could add
> > something like variabes = {"a": 2.0, "b": 3.0}.
>
> No, I think it is nice to have them as **kwargs.
>
> My point is that these arguments need to be last in a function signature. That
> again means that they need to come after default kwargs like element, cell,
> degree, forcing the syntax of the first example above.
>
> However it is more intutive to keep initialization of these argument closest
> to the string, the second syntax example. But then we need to conceal valid
> kwargs into an anonymous *kwargs.
>
> That said, I *really* think it is bad habit to conceal valid keyword arguments
> into **kwargs. I cannot tell how many times I have hit:
>
>   plot?
>
> in ipython just to get dissapointed of what I need to feed it.
>
> Johan

ok. This would be a nice feature to have. Add it if/when you feel like
it, but it might be good to have it in place before the release of 1.0
since it means a change in the interface.

--
Anders