example: custom numeric type

[ofloveandhate]

This pull request is an example of building against EigenPy to expose Eigen objects with custom numeric types from C++. This is in response to #365.

This example at time of PR builds and installs, but not fully correctly. It doesn't install usable Python bindings, so the example Python code is not runnable. Maybe you could do that last step? So that the built bindings can actually be used in Python?

[jcarpent]

Thanks @ofloveandhate for the PR. I will make a pass on it when I have time in order to align it with the rest of the project. Is it fine for you?

[jcarpent]

TODO:

• Remove useless files
• SImplify cmake usage
• Extend Python test with the creation of Numpy arrays used by a C++ program

[ofloveandhate]

I tried to use the code in this PR, and it doesn't install correctly on my Mac. I observe the following issues:

• I needed to change line 3 in CMakeLists.txt to cmake_minimum_required(VERSION 3.10)
• The built library, when installed, has extension .dylib instead of .so
• The build library's name starts with lib, but it needs to just start with the actual name of the library.
• If imported, I get missing symbols:

import libeigenpy_example_custom_numeric_type

ImportError: dlopen(/usr/local/lib/libeigenpy_example_custom_numeric_type.so, 0x0002): symbol not found in flat namespace '_EIGENPY_ARRAY_API'

I think all of these are issues with the CMakeLists.txt file.

[ofloveandhate]

I would love to be able to build upon this to work on some problems I'm seeing with custom numeric types in EigenPy, and make this example more complete, but I am finding the setup stuff is a barrier. Thanks!

[ofloveandhate]

I'm back for more help, and to help this example of using Eigenpy with a custom numeric type become fully functional.

I just added some code that makes EigenPy / the example library crash. I think my code is missing something simple. The failing code in Python is merely this:

import eigenpy_example_custom_numeric_type as example
import numpy as np

M = np.zeros((3,4),dtype=example.MpfrComplex)  # make an array of the custom numeric type

The problem is clearly uninitialized data, as per the assertion fail from MPC:

Assertion failed: (m_data[0].re[0]._mpfr_d), function data, file mpc.hpp, line 346.

(This is a step towards another problem I have with alignment)

---

I would love some help so that I can make arrays with my custom type in Numpy, and then feed them to C++ routines. I think this PR is a good place to do it, so that the resulting example code is contributed directly to EigenPy. (I am sorry it took me so long to follow up -- I switched jobs, moved continents, and am now in a full time research position 🥳)

[ofloveandhate]

I'm getting closer to the problem I'm trying to identify:

how to be able to make numpy containers of custom numeric type, via EigenPy, without converting from existing.

This most recent set of commits shows that I can make numpy array of custom dtype (MpfrFloat and MpfrComplex) IF I FIRST make an array in built-in type, and convert to my own type.

A = np.array( np.empty( (3,4)).astype(np.int64),dtype=num_type) 
# this is used in the user_type example in EigenPy

But, the following still fails, which feels like a showstopper:

B = np.zeros( (4,5), dtype=num_type)

What magic sauce am I still missing to be able to directly make numpy arrays of custom type?