← Back to team overview

dolfin team mailing list archive

Re: Problem with compile_function

 

On Saturday 06 December 2008 14:48:34 Anders Logg wrote:
> On Sat, Dec 06, 2008 at 02:41:55PM +0100, Johan Hake wrote:
> > On Saturday 06 December 2008 14:23:40 Anders Logg wrote:
> > > On Sat, Dec 06, 2008 at 12:02:43PM +0100, Johan Hake wrote:
> > > > On Friday 05 December 2008 18:02:07 Anders Logg wrote:
> > > > > I'm having problems getting vector-valued constants working. Take a
> > > > > look at
> > > > >
> > > > >     demo/pde/stokes/taylor-hood/python
> > > > >
> > > > > It reports
> > > > >
> > > > >     assert(isinstance(defaults[i], (dict, types.NoneType)),"Wrong
> > > > > type in 'defaults'")
> > > > >    TypeError: 'NoneType' object is unsubscriptable
> > > > >
> > > > > It seems like defaults is set to None and then defaults is indexed.
> > > > >
> > > > > Any ideas?
> > > >
> > > > Should be fixed now. I have added better argument checking to
> > > > compile_function. But you have to send in a tuple of strings to
> > > > compile_function if you want to produce a vector values function. A
> > > > list is interpreted as len(cppexpr) number of scalar functions.
> > > >
> > > > This is a bit fragile, but it is documented in the Function doc
> > > > string.
> > >
> > > ok! I wasn't aware of the difference.
> > >
> > > Maybe we should have both compile_function and compile_functions
> > > instead of differentiating between tuples and lists?
> > >
> > > There is already a compile_functions defined in compile_functions.py,
> > > but that does not seem to be used. Should it be removed (and the name
> > > reused for batch-processing functions)?
> >
> > compile_functions.py is the orginal work of Martin. It is depricated now.
> >
> > The compile_function module should not be exposed to the user, (Martin
> > might have a different opinion here) so it should be sufficient with the
> > Function interface.
> >
> > The differences between compiling one and several functions are so small
> > that I do not think it is usefull to split the code into different
> > modules.
> >
> > One could use the name Functions, instead of Function for batch
> > processing. In this way we make clear for a user what it means. This also
> > correlates with the TestFunction/TestFunctions.
> >
> > Unfortunaltly will not the case of defining several functions in a
> > cppcode, argument be covered by this, as the number of compiled functions
> > will depend on how many functions the user defines in the cppcode. For
> > this case we could add CostumFunctions, or something.
> >
> > Or just keep it the way it is with clearer documentation?
>
> Wouldn't it be easy to add an additional argument batch=False to
> compile_function? Then you could check that argument instead of
> differentiating between lists and tuples.

So,

  f = Function(V,cppexpr = (("sin(x[0])","cos(x[1])"),
                           (("sin(x[1])","exp(x[1])"))))

would create a tensor valued function and:

  f, g = Function(V,cppexpr = (("sin(x[0])","cos(x[1])"),
                              (("sin(x[1])","exp(x[1])")),
                               batch = True)

would create two vector valued functions. Then we need to explain that the use 
of batch = True, change the meaning of the outermost parentesis from a matrix 
expression to a list expression of two vectors. I have bad feeling...

We could also be a bit radical and adapt a pylab.plot interface to the 
compile_function module 

  f, c = FunctionExpression(V0,("sin(x[0])","cos(x[1])"),V1,"2.5")

where a space followed by an expression will associate the space with that 
expression. We can have an optional number of (V, cppexpr) combinations. And 
the natural extension of:

  f0, f1, f2  = FunctionExpression("sin(x[0])","cos(x[1])","2.5", space = V)

where each compiled function is instantiated with the function space defined 
in 'space'. This could be followed up with e.g:

  CostumFunction(V0,code0,V1,code1)

and

  CostumFunction(code0,code1, space = V)

(The names may not be the best...)

Johan


Follow ups

References