Qeustion using @bempp.api.complex_callable

[Sangwoo Kang]

Dear all

Hello. I try to translate my bempp code as bempp-cl to solve large problem in windows. 

But I got the error messages when I define the Gridfunction . 

My incident field is generated by surface currents J_DUT, M_DUT. 

It is working in the bempp 3.3.4 version with docker. 

But the following code is not working in bempp-cl when I define the IncidentEfiled function as the following: 

case1. 

def IncidentEfield(point):

    temppoint = point.reshape((3,1))

    PropLpart = bempp.api.operators.potential.maxwell.electric_field(div_space_DUT, temppoint,k0)

    PropKpart = bempp.api.operators.potential.maxwell.magnetic_field(div_space_DUT, temppoint,k0)

    JJ = J_DUT*eta0

    MM = M_DUT

    tempresult = (PropLpart*JJ - PropKpart*MM)

    value = tempresult.reshape(-1)

    return value

@bempp.api.complex_callable

def tangential_trace(point, n, domain_index, result):

    value = IncidentEfield(point)

    result[:] = np.cross(value, n)

I got a error message TypingError: cannot determine Numba type of <class 'function'>.

case2. 

@numba.jit

def IncidentEfield(point):

    temppoint = point.reshape((3,1))

    PropLpart = bempp.api.operators.potential.maxwell.electric_field(div_space_DUT, temppoint,k0)

    PropKpart = bempp.api.operators.potential.maxwell.magnetic_field(div_space_DUT, temppoint,k0)

    JJ = J_DUT*eta0

    MM = M_DUT

    tempresult = (PropLpart*JJ - PropKpart*MM)

    value = tempresult.reshape(-1)

    return value

@bempp.api.complex_callable

def tangential_trace(point, n, domain_index, result):

    value = IncidentEfield(point)

    result[:] = np.cross(value, n)

In this case, I got a message 

TypingError: Invalid use of type(CPUDispatcher(<function IncidentEfield at 0x000001E6EE592DC8>)) with parameters (array(float64, 1d, A))

 * parameterized During: resolving callee type: type(CPUDispatcher(<function IncidentEfield at 0x000001E6EE592DC8>))

Please let me know if you have any idea to handle this issue. 

Thank you in advance. 

BR,

Sangwoo Kang

[Matthew Scroggs]

Hello,

Numba can's compile your function, and I think it would be very difficult or even impossible to rewrite it in a way that numba could compile it.

Currently, Bempp-cl always uses Numba when assembling a GridFunction from a callable. I've added a pull request (https://github.com/bempp/bempp-cl/pull/17) to allow the user to change this. Once this is merged (or if you check out the mscroggs/no_jit branch), you should be able to make your GridFunction using:

bempp.api.GridFunction(space, fun=tangential_trace, jit=False, function_type='complex')

[Timo Betcke]

Hi,

there is a way around this. You can use the  objmode context manager within your grid function code to denote a block that should not be compiled with Numba.

A description is here: https://numba.pydata.org/numba-doc/dev/user/withobjmode.html

However, be warned. This slows down the grid function generation significantly as the corresponding code now needs to go through the Python interpreter.

Best wishes

Timo