dolfin team mailing list archive

[A Navaei <axnavaei@xxxxxxxxxxxxxx>]

2009/2/14 Johan Hake <hake@xxxxxxxxx>:
> On Saturday 14 February 2009 13:48:05 A Navaei wrote:
>> It seems that the error can be re-produced even without sub-classing
>> and using existing dolfin classes in pure c++. Based on the itk-dolfin
>> interface code, the below minimal code should generate the error (note
>> that I use the binary distribution which uses std::tr1, replacing it
>> with boost shared pointer should not have any effect).
>
> Not in c++ but swig only support std::tr1::shared_ptr from version 1.3.37. But
> if you intend to use shared_ptr only internally there whould not be any
> problems.
>
>> After wrapping in python:
>>
>> (1) Calling FunctionTest.CreateFunction(), which returns by value,
>> results this error:
>>
>> RuntimeError: *** Error: Unable to assign to function, missing
>> coefficients (user-defined function).
>
> You cannot assign another user-defined function to another Function. It must
> be a discrete function, which has an initialized _vector. This is probably a
> feature that other developers should answer for.
>
> However this means that you cannot copy a userdefined function, with the side
> effect of not beeing able to return a user-defined Function by value.
>
>> Since dolfin::Function does come with the required copy ctors, the
>> problem cannot be stemmed from this.
>
> This use the assignment operator which requires the Function to be a discrete
> and not a user-defined Function.
>
> I consider this to be a bug in the library. Any other comments from the C++
> DOLFIN developers (I am mostly dealing with the python interface)?
>
>> (2) Calling FunctionTest.CreateFunctionPtr(), which returns the shared
>> pointer, does not generate any errors.
>
> This should be expected, as the copy constructors is not called. When you get
> all this to work, eventually ;), I would suggest using shared_ptr types for
> the return argument, as shared_ptrs are much nicer to deal with in both c++
> and python. But then you probably need the development version of DOLFIN. We
> do consider a release soon, which you can switch to when this is out.
>
> The support for shared_ptr in PyDOLFIN is lately added, but has more or less
> stabilized now, I think.
>
>> However, there are problems
>> with handling the shared pointer in python. I used this in my swig
>> interface file (or in an implicit way you can include dolfin.i
>> defining the right flags):
>>
>> #define SWIG_SHARED_PTR_NAMESPACE std // comment out if it's boost
>> #define SWIG_SHARED_PTR_SUBNAMESPACE tr1 // comment out if it's boost
>> %include "boost_shared_ptr.i"
>> # if defined(SWIG_SHARED_PTR_QNAMESPACE)
>> SWIG_SHARED_PTR(Function, dolfin::Function)
>> #endif
>
> Which version of swig do you have, refering to the comment above. You need
> swig version >= 1.3.35 to be able to use shared_ptr in the first place and >=
> 1.3.37 to be able to use std::str1::shared_ptr.

I'm using swig 1.3.38, I noticed that the patch you submitted to swig
was used in 1.3.37; but the above does not work for me. I tried using
the swig interface code, before/after the class declaration and still
shared_ptr does not get wrapped in the desired way. Even following
this http://article.gmane.org/gmane.comp.programming.swig/12803 didn't
help.

I noticed that, at least for the latest binary release, shared
pointers are not wrapped properly in python. Try this:

mesh = UnitSquare(100, 100)
V = FunctionSpace(mesh, "CG", 1)
f = Function(V)
f.function_space_ptr()

Out[17]: <Swig Object of type 'std::tr1::shared_ptr<
dolfin::FunctionSpace const > *' at 0x9728b80>

I guess the above returned object is useless? That's what I get for
returned Function in my code.

>
>> Having this, in python, FunctionTest.CreateFunctionPtr() returns a
>> proxy to the shared pointer object which is not useful. Are there any
>> steps missed in the wrapping process?
>
> Strictly speaking it does _not_ return a proxy but rather a raw pointer to the
> object. The proxy is the python wrapper class that swig creates for us when
> it have the correct type knowledge, which in this case, swig does not have.
>
>> c++ test code:
>> -----------------------------
>> class FunctionTest
>> {
>> public:
>>       typedef std::tr1::shared_ptr<const dolfin::Mesh> MeshConstPointerType;
>>       typedef std::tr1::shared_ptr<const dolfin::FiniteElement>
>> ElementConstPointerType;
>>       typedef std::tr1::shared_ptr<const dolfin::DofMap> DofMapConstPointerType;
>>       typedef std::tr1::shared_ptr<const dolfin::FunctionSpace>
>> FSConstPointerType; typedef std::tr1::shared_ptr<dolfin::Function>
>> FPointerType;
>>
>>       static dolfin::Function CreateFunction()
>>       {
>>               dolfin::Mesh mesh = dolfin::UnitSquare(100, 100);
>>               std::string elemSig("FiniteElement('Lagrange', 'triangle', 1)");
>>               std::string dofSig("FFC dof map for FiniteElement('Lagrange',
>> 'triangle', 1)");
>>
>>               FSConstPointerType fs(new dolfin::FunctionSpace(
>>               typename IFSType::MeshConstPointerType(&mesh,
>> dolfin::NoDeleter<const dolfin::Mesh>()),
>>               typename IFSType::ElementConstPointerType(new
>> dolfin::FiniteElement(elemSig)), typename
>> IFSType::DofMapConstPointerType(new dolfin::DofMap(dofSig, mesh))) );
>>
>>               dolfin::Function func(fs);
>>               return func;
>>       };
>>
>>       static FPointerType CreateFunctionPtr()
>>       {
>>               dolfin::Mesh mesh = dolfin::UnitSquare(100, 100);
>>               std::string elemSig("FiniteElement('Lagrange', 'triangle', 1)");
>>               std::string dofSig("FFC dof map for FiniteElement('Lagrange',
>> 'triangle', 1)");
>>
>>               FSConstPointerType fs(new dolfin::FunctionSpace(
>>               typename IFSType::MeshConstPointerType(&mesh,
>> dolfin::NoDeleter<const dolfin::Mesh>()),
>>               typename IFSType::ElementConstPointerType(new
>> dolfin::FiniteElement(elemSig)), typename
>> IFSType::DofMapConstPointerType(new dolfin::DofMap(dofSig, mesh))) );
>>
>>               FPointerType funcp = FPointerType(new dolfin::Function(fs));
>>               return funcp;
>>       };
>> };
>> ----------------------------
>> > Yes, I can call wrapped DolfinImageFunction in python with no problem,
>> > but when using it in pure c++ (itk::ImageToDolfinFunction) an then
>> > calling ImageToDolfinFunction after wrapped in python generated the
>> > error:
>> >
>> > RuntimeError: *** Error: Unable to assign to function, missing
>> > coefficients (user-defined function).
>
> I think we sorted out this problem above: do not return by value.
>
>> > why v._vector is not null in the first case? I also tried the shared
>> > pointer ctors and ended up with the same result. Moreover, I tried
>> > sub-classing FunctionSpace (see
>> > http://code.google.com/p/wrapitk/source/browse/trunk/ExternalProjects/Itk
>> >Dolfin/src/itkDolfinImageFunctionSpace.h), with and without shared pointer
>> > ctors, again the same error.
>> >
>> >> First I assume that you properly instantiate the templates somewhere
>> >> else in your build script, as itkDolfin.swg does not define any such
>> >> templates.
>> >
>> > That's correct, wrapitk instantiates the templates automatically.
>
> Ok. Is this done in some of the SMake macros? I am really not familiar to
> CMake

No, it's done in wrapitk using cableswig (which itself uses gccxml),
python and cmake. The end user simply specifies the desired wrapping
types. In this thread, let's assume that automatic template
instantiation is properly done.

>
>> >> Second you should probably return a pointer to the created function in
>> >> DolfinImageFunction(), and set
>> >>
>> >>  %newobject
>> >> itk::ImageToDolfinFunction<YourImageTypes>.DolfinImageFunction;
>> >>
>> >> or define a copy constructor to your DolfinImageFunction, as you are
>> >> returning by value now.
>> >
>> > I've tried a simple copy ctor like this:
>> >
>> >        DolfinImageFunction(const DolfinImageFunction &v)
>> >        {
>> >                 *this = v;
>> >        }
>> >
>> > which does get called, but didn't help.
>
> Again, bug in copy constructor. You should be able to return a raw pointer to
> the function, but make sure you use the %newobject directive mentioned above,
> to till swig that it should take ownership to the returned object.

The template instantiator defines a class like:

itkImageToDolfinFunctionID2

which stands for an image of dimention 2 and type double. Given that
the shared pointer object is created in Convert() method, shouldn't
the memory leak problem be solved by something like:

%newobject itkImageToDolfinFunctionID2::Convert;

It doesn't work for me. At least this should work:

%newobject ::Convert;

which should cover every function called 'Convert' in any classes, but
it dosn't work either. I also tried this before/after class
declaration. Where is the similar code in dolfin?

>
>> > Just to mention that one problem with c++ dolfin classes is that their
>> > member variables are mostly private, causing restrictions on the
>> > sub-classes.
>
> This should others answer.
>
>> > Is there a specific reason for having private variables
>> > instead of protected? For instance, DolfinImageFunction ctors are not
>> > able to initialise _function_space or _vector.
>
> This should not be any problems in the ordinary constructor as you can just
> pass the function_space to the super constructor, and a properly working copy
> constructor should also fix this for any derived copy constructors.

In the ImageFunction:Function case, the ctor takes an argument of
image type, while the parent class arguments objects (Mesh,
FiniteElement and DofMap) are required to be initialised based on the
image object. With private member variables, if not impossible, it is
difficult to do so. My work-around was sub-classing from FunctionSpace
and assigning it to the Function object later.

>
> The _vector is created by calling init(). But you do not want to do this as
> the Function will then change status from a user-defined to a discrete
> function, making it impossible to reach the eval function.

init() is not called in any of the copy ctors of Function, this could
be a source of the bug.

>
> Again hope this helps.

Thanks for the detailed comments.

-Ali

>
> Johan
>