Very odd Core Dump when importing numpy while exe is linked with libSDL2.so

[GavinNL]

I'm having a very odd issue when trying to embed the interpreter into my C++ program and cannot seem to find any kind of solution. I'm hoping someone here can help.

I have an OpenGL application, which links to three shared libs that I wrote, libA, libB, and libSDLWindow
All are compiled as shared objects. libSDLWindow also privately links to libSDL2.so

I'm trying to embed the interpreter into my C++ code, but I'm getting odd behaviour when I
link to libSDLWindow.so. (this is just shared lib that helps me create SDL Windows for rendering)

In the code below, if I do not include auto sdlWindow = std::make_shared<SDLWindow>();, gcc wont link
libSDL2.so to the final exe, and the python code executes properly.

If I keep the code in and libSDL2.so is linked with my application, the py::exec line will core dump.
This only happens when i'm importing the numpy library. If I remove the numpy import, everything works fine

I'm not sure why libSDL2.so is somehow conflicting with the import numpy.

Some Notes:

• The core dump occurs on the py::exec call not on the auto sdlWindow = std::make_shared<SDLWindow>();
• I'm using python 3.6 on Ubuntu 18.04. Numpy is installed via pip3 and is located in the user's home directory.
• I don't think the problem is with numpy, it seems to be any python lib that's not the standard libraries
• Standard libraries seem to import fine.
• libA and libB do not seem to cause any issues, just the sdlWindow library

#include <libA.h>
#include <libB.h>
#include <SDLWindow.h> // built as shared lib, links to libSDL2.so

#include <pybind11/pybind11.h>
#include <pybind11/embed.h>

namespace py = pybind11;


int main(int argc, char ** argv)
{
    py::initialize_interpreter();

    py::exec(R"(
             import numpy
             import sys
             kwargs = dict(name="World", number=42)
             message = "Hello, {name}! The answer is {number}".format(**kwargs)
             print(message)
             print(sys.path)
             print(numpy.array([1,2,3,4]))
             )");

    // Comment this line, and py::exec will work
    // leave it in, py::exec will abort with (core dumped)
    //
    // not calling py::exec will not cause a crash
    auto sdlWindow  = std::make_shared<SDLWindow>();


    auto A      = std::make_shared<classA>();
    auto B     = std::make_shared<classB>();

    py::finalize_interpreter();
    return 0;
}