how to pythonize a NumProc? My-Little-Ngsolve question

Hello,

I have trouble using NumProc in python.
More specifically, in the file myAssembling.cpp, we use
static RegisterNumProc<...> ..("myassemble")
Then, the pde file recognize myassemble as a numproc, and we can call it using the command
numproc myassemble np1

The question is how to call this NumProc in python. It seems I need to register this numproc in python to use it... Is this related to "pythonize the code"?

I dig a bit into the source code but find nothing.... What I noticed is that the names created in RegisterFESpace can be used in <FESpace>, the names created in RegisterBilinearFormIntegrator can be used in <BFI>, similar for <LFI>, and the name created in RegisterPreconditioner can be used in <Preconditioner>. So, naively, I thought the name created by RegisterNumProc shall be used in <NumProc>, but this is not the case...

Best,
Guosheng

Hi Guosheng,

short answer: The Pythonic way is not using NumProcs anymore. Whatever you do in the NumProc::Do() working function, you can do in a Python function. The old PDE class contains a list of NumProcs, which are processed in order. This is not needed anymore, since the program-flow is now steered from Python. As you mentioned myassemble, this can be done perfectly from Python.
It makes sense to register FESpaces from C++ and access them in Python. Here a lot of functionality is programmed in C++, and there is a small common interface. The same holds for preconditioners.
Specific LinearForm/BilinearForm-integrators from C++ should not be needed anymore, one should be able to define the form using SymbolicBFI/SymbolicLFI.

Of course, it will be important also in future to export C++ functions to Python, to perform some specific functionality more efficient and in parallel. Here you have to Python-wrap some functions (or also classes).

Practically, for the transition time from old PDE-style to new Python-style, we also wrapped NumProcs to Python, such that existing NumProcs could be used. See the example of the "BVP" numproc (file solve/bvp.cpp). Here we exported a creator-function (line 789) to create a NumProcBVP from Python.

See also mylittlengsolve/python/myngspy.cpp showing how to write extensions in extra shared libraries (comment: I just updated from boost::python to pybind11, needs some more testing).

Hope this explains the ideas, if there are particular questions left, let us know.

Best, Joachim