Investigate runtime savings of simplifying the code we swig when jitting ufc forms

[Martin Sandve Alnæs]

If we can get away with generating a factory function for a ufc::form and just swigging a header with the signature, and hiding all the generated code in a .cpp file, we might be able to reduce the overhead of the jit compilation significantly. This should of course be investigated before doing surgery on ffc.

I.e. we can generate

// form_<hashvalue>.h
#include <ufc.h>
ufc::form * create_form_<hashvalue>();

// form_<hashvalue>.cpp
#include "form_<hashvalue>.h"

... regular generated ufc code
class form_<hashvalue>: public ufc::form
{ ... };

ufc::form * create_form_<hashvalue>()
{
    return new form_<hashvalue>();
}

[Martin Sandve Alnæs]

We could possibly even avoid swig altogether for this simplified generated code by returning a void* and using ctypes.

// form_<hashvalue>.h
void * create_form_<hashvalue>();

// form_<hashvalue>.cpp
#include <ufc.h>
#include "form_<hashvalue>.h"

... regular generated ufc code
class form_<hashvalue>: public ufc::form
{ ... };

void * create_form_<hashvalue>()
{
    return new form_<hashvalue>();
}

[Prof Garth Wells]

I think we should remove the FFC SWIG dependency, and use ctypes.

[Martin Sandve Alnæs]

I implemented a ctypes based instant replacement a week ago. It's in a private sandbox repo, I'll clean it up and publish it soon. It also has some mpi awareness with mpi4py.

[Martin Sandve Alnæs]

Investigation is done. This is definitely valuable, and other issues for doing the work has been created.

[Prof Garth Wells]

Removing milestone: 1.7 (automated comment)