I am currently trying to wrap a c-dll to controll a camera connected via USB. To grab image data off the camera, the SDK-Provider (Thorlabs) describes two possibilities:
Poll the camera in a loop
Use callbacks
I wrapped the DLL and got the polling method to work, which basically consists of calling a method which returns either NULL or the pending frame.
The DLL also provides methods to assign callback functions to certain events, such as camera connected/disconnected and frame available. The C-code from the example looks like this:
void camera_connect_callback(char* cameraSerialNumber, enum USB_PORT_TYPE usb_bus_speed, void* context)
{
printf("camera %s connected with bus speed = %d!\n", cameraSerialNumber, usb_bus_speed);
}
int main(void)
{
[...]
// Set the camera connect event callback. This is used to register for run time camera connect events.
if (tl_camera_set_camera_connect_callback(camera_connect_callback, 0))
{
printf("Failed to set camera connect callback!\n");
close_sdk_dll();
return 1;
}
[...]
}
My goal is now to get the callback working with python, so that python methods are called from the dll. Therefore, I used CFFI in API-mode to compile the wrapper, then tried to pass a callback method to the DLL in the same fashion:
#ffi.callback("void(char*, void*)")
def disconnect_callback(serialNumber, context):
print("Disconnect callback was called!")
[...]
print("Set disconnect callback: {}".format(lib.tl_camera_set_camera_disconnect_callback(disconnect_callback, ffi.NULL)))
[...]
For testing purposes I only included the simplest callback in the working polling example. The setter method returns 0 but the method is never executed.
Is this even the right way of accomplishing my goal? Do I need to introduce some kind of threading so that the callback can interrupt the sequential execution of the remaining program?
There is not much documentation or examples on this topic so Im really hoping you guys can help me out.
EDIT
So I tried a very basic example which works and technically answeres my question, but does not solve my problem.
if I define a C callback method as follows:
#include <math.h>
int quadcalc(int input, int (*getval)(int)){
return (int) pow(getval(input), 2);
}
I can assign a handler as expected: (After compiling of course)
#ffi.callback("int(int)")
def callback_func(value):
return int(rnd.random() * value)
if __name__ == '__main__':
print(lib.quadcalc(10, callback_func))
And everything works as expected. Unfortunately, this doesn't work on the actual Problem. So callback works in general, but not in the specific case, which is why the question is still open.
Do I need to introduce some kind of threading so that the callback can interrupt the sequential execution of the remaining program?
Invoking the Callback method takes care of setting up alternate program flows to capture the event, so it can be routed to a handler to be processed. When the defined event occurs, it does need handler code. For example, here is a very generic, but complete example of code that uses a callback in C (including the handler function.):
void populate_array(int *array, size_t arraySize, int (*getNextValue)(void))
{
for (size_t i=0; i<arraySize; i++)
array[i] = getNextValue();
}
// handler function
int getNextRandomValue(void)
{
return rand();
}
int main(void)
{
int myarray[10];
populate_array(myarray, 10, getNextRandomValue);
...
}
There is an equally good example just below the first one, here.
Beyond these, there is a good tutorial in the CFFI documentation, with the first section dedicated to addressing your title question How to make a C-DLL wrapped with CFFI to callback python:
The first section presents a simple working example of using CFFI to call a C function in a compiled shared object (DLL) from Python.
Steps are: (see link above for details of each step.)
- Create the file piapprox_build.py:
- Execute this script:
- At runtime, you use the extension module like this:
...In the rest of this page, we describe some more
advanced examples and other CFFI modes...
Related
My python code has been crashing with error 'GC Object already Tracked' . Trying to figure out the best approach to debug this crashes.
OS : Linux.
Is there a proper way to debug this issue.
There were couple of suggestions in the following article.
Python memory debugging with GDB
Not sure which approach worked for the author.
Is there a way to generate memory dumps in such scenario which could be analyzed. Like in Windows world.
Found some article on this. But not entirely answers my question:
http://pfigue.github.io/blog/2012/12/28/where-is-my-core-dump-archlinux/
Found out the reason for this issue in my scenario (not necessarily the only reason for the GC object crash).
I used the GDB and Core dumps to debug this issue.
I have Python and C Extension Code (in shared object).
Python code registers a Callback routine with C Extension code.
In a certain workflow a thread from C Extension code was calling the registered Call back routine in Python code.
This usually worked fine but when multiple threads did the same action concurrently it resulted in the Crash with 'GC Object already tracked'.
Synchronizing the access to python objects for multiple thread does resolve this issue.
Thanks to any responded to this.
I ran into this problem using boost::python when our C++ code would trigger a python callback. I would occasionally get "GC object already tracked" and the program would terminate.
I was able to attach GDB to the process prior to triggering the error. One interesting thing, in the python code we were wrapping the callback with
a functools partial, which was actually masking where the real error was occuring. After replacing the partial with a simple callable wrapper class. The "GC object already tracked error" no longer popped up, instead I was now just getting a segfault.
In our boost::python wrapper, we had lambda functions to handle a C++ callback and the lambda function captured the boost::python::object callback function. It turned out, for whatever reason, in the destructor for the lambda, it wasn't always properly acquiring the GIL when destroying the boost::python::object which was causing the segfault.
The fix was to not use a lambda function, but instead create a functor that makes sure to acquire the GIL in the destructor prior to calling PyDECREF() on the boost::python::object.
class callback_wrapper
{
public:
callback_wrapper(object cb): _cb(cb), _destroyed(false) {
}
callback_wrapper(const callback_wrapper& other) {
_destroyed = other._destroyed;
Py_INCREF(other._cb.ptr());
_cb = other._cb;
}
~callback_wrapper() {
std::lock_guard<std::recursive_mutex> guard(_mutex);
PyGILState_STATE state = PyGILState_Ensure();
Py_DECREF(_cb.ptr());
PyGILState_Release(state);
_destroyed = true;
}
void operator ()(topic_ptr topic) {
std::lock_guard<std::recursive_mutex> guard(_mutex);
if(_destroyed) {
return;
}
PyGILState_STATE state = PyGILState_Ensure();
try {
_cb(topic);
}
catch(error_already_set) { PyErr_Print(); }
PyGILState_Release(state);
}
object _cb;
std::recursive_mutex _mutex;
bool _destroyed;
};
The problem is that you try to add an object to Python's cyclic garbage collector tracking twice.
Check out this bug, specifically:
The documentation for supporting cyclic garbage collection in Python
My documentation patch for the issue and
My explanation in the bug report itself
Long story short: if you set Py_TPFLAGS_HAVE_GC and you are using Python's built-in memory allocation (standard tp_alloc/tp_free), you don't ever have to manually call PyObject_GC_Track() or PyObject_GC_UnTrack(). Python handles it all behind your back.
Unfortunately, this is not documented very well, at the moment. Once you've fixed the issue, feel free to chime in on the bug report (linked above) about better documentation of this behavior.
I've just started using OpenCv 3.1 and have encountered the following annoying behavior. Whenever I make an initial call to imshow (actually cv2.imshow, since I'm using the Python interface), I get this output to my screen:
init done
opengl support available
It seems to be due to the following method in window_QT.cpp:
static int icvInitSystem(int* c, char** v)
{
//"For any GUI application using Qt, there is precisely one QApplication object"
if (!QApplication::instance())
{
new QApplication(*c, v);
setlocale(LC_NUMERIC,"C");
qDebug() << "init done";
#ifdef HAVE_QT_OPENGL
qDebug() << "opengl support available";
#endif
}
return 0;
}
All I can think of to do is to comment out the qDebug calls and recompile OpenCV. Is there any less drastic solution that would either automatically redirect qDebug's output to stderr, or just turnoff debug information unless I actively want it?
qDebug is a preprocessor-controlled, but it has its own special macro, QT_NO_DEBUG_OUTPUT. If you add that to your Release build defines, it will be removed.
Im building a C++ library which gets called by python via ctypes. The function has as parameter a pointer to a struct and passes it to other (internal) function. It looks like this:
extern "C" __declspec(dllexport) void USBEndpoint_subscribe(USBEndpoint* endpoint, CALLBACK callback) {
try {
subscribe_internal(endpoint, callback);
} catch (int e_code) {
exception_report(__FILE__, __LINE__, e_code); \
} catch (...) {
exception_report(__FILE__, __LINE__);
}
void subscribe_internal(USBEndpoint* endpoint, CALLBACK callback)
{
...
}
The thing is that the value for endpoint recieved at USBEndpoint_subscribe is not the same value that enters subscribe_internal. For example:
while debugging I can see that USBEndpoint_subscribe recieves:
endpoint=0x00000000088f0680 (and callback=0x0000000000540f50)
and as soon as I step into subscribe_internal, inside subscribe_internal the values are:
endpoint=0x000007fef15f7740 (and callback=0x0000000000000000).
I have not spawn other threads and I'm compiling the library in release mode**. My only suspect could be stack corruption, but actually I have clue of what's going on.
Any hints of what can be happening are very welcome.
** I have compiled python in debug mode, but it's slow to set up with spider and I loose the interactive console if I use my debug-compiled version, so I prefer sticking with release python and compiling my library in release.
EDIT:
As suggested, I used my debug version of python with a debug compile of the library. But I still got the same behavior (endpoint: 0x0000000002be80e0->0x0000000000000000, callback: 0x0000000000320f88->0x00000000024d24b8). No other useful information appeared here that wasn't in release mode.
I was loading The library using CDLL("path"), which makes the calling convention cdecl. And even if I'm compiling in VS2008 with /Gd, it seems it's using other convention. Appending __cdecl to the function declaration solves the issue.
EDIT: Ok, all the edits made the layout of the question a bit confusing so I will try to rewrite the question (not changing the content, but improving its structure).
The issue in short
I have an openCL program that works fine, if I compile it as an executable. Now I try to make it callable from Python using boost.python. However, as soon as I exit Python (after importing my module), python crashes.
The reason seems to have something to do with
statically storing only GPU CommandQueues and their release mechanism when the program terminates
MWE and setup
Setup
IDE used: Visual Studio 2015
OS used: Windows 7 64bit
Python version: 3.5
AMD OpenCL APP 3.0 headers
cl2.hpp directly from Khronos as suggested here: empty openCL program throws deprecation warning
Also I have an Intel CPU with integrated graphics hardware and no other dedicated graphics card
I use version 1.60 of the boost library compiled as 64-bit versions
The boost dll I use is called: boost_python-vc140-mt-1_60.dll
The openCL program without python works fine
The python module without openCL works fine
MWE
#include <vector>
#define CL_HPP_ENABLE_EXCEPTIONS
#define CL_HPP_TARGET_OPENCL_VERSION 200
#define CL_HPP_MINIMUM_OPENCL_VERSION 200 // I have the same issue for 100 and 110
#include "cl2.hpp"
#include <boost/python.hpp>
using namespace std;
class TestClass
{
private:
std::vector<cl::CommandQueue> queues;
TestClass();
public:
static const TestClass& getInstance()
{
static TestClass instance;
return instance;
}
};
TestClass::TestClass()
{
std::vector<cl::Device> devices;
vector<cl::Platform> platforms;
cl::Platform::get(&platforms);
//remove non 2.0 platforms (as suggested by doqtor)
platforms.erase(
std::remove_if(platforms.begin(), platforms.end(),
[](const cl::Platform& platform)
{
int v = cl::detail::getPlatformVersion(platform());
short version_major = v >> 16;
return !(version_major >= 2);
}),
platforms.end());
//Get all available GPUs
for (const cl::Platform& pl : platforms)
{
vector<cl::Device> plDevices;
try {
pl.getDevices(CL_DEVICE_TYPE_GPU, &plDevices);
}
catch (cl::Error&)
{
// Doesn't matter. No GPU is available on the current machine for
// this platform. Just check afterwards, that you have at least one
// device
continue;
}
devices.insert(end(devices), begin(plDevices), end(plDevices));
}
cl::Context context(devices[0]);
cl::CommandQueue queue(context, devices[0]);
queues.push_back(queue);
}
int main()
{
TestClass::getInstance();
return 0;
}
BOOST_PYTHON_MODULE(FrameWork)
{
TestClass::getInstance();
}
Calling program
So after compiling the program as a dll I start python and run the following program
import FrameWork
exit()
While the import works without issues, python crashes on exit(). So I click on debug and Visual Studio tells me there was an exception in the following code section (in cl2.hpp):
template <>
struct ReferenceHandler<cl_command_queue>
{
static cl_int retain(cl_command_queue queue)
{ return ::clRetainCommandQueue(queue); }
static cl_int release(cl_command_queue queue) // -- HERE --
{ return ::clReleaseCommandQueue(queue); }
};
If you compile the above code instead as a simple executable, it works without issues. Also the code works if one of the following is true:
CL_DEVICE_TYPE_GPU is replaced by CL_DEVICE_TYPE_ALL
the line queues.push_back(queue) is removed
Question
So what could be the reason for this and what are possible solutions? I suspect it has something to do with the fact that my testclass is static, but since it works with the executable I am at a loss what is causing it.
I came across similar problem in the past.
clRetain* functions are supported from OpenCL1.2.
When getting devices for the first GPU platform (platforms[0].getDevices(...) for CL_DEVICE_TYPE_GPU) in your case it must happen to be a platform pre OpenCL1.2 hence you get a crash. When getting devices of any type (GPU/CPU/...) your first platform changes to be a OpenCL1.2+ and everything is fine.
To fix the problem set:
#define CL_HPP_MINIMUM_OPENCL_VERSION 110
This will ensure calls to clRetain* aren't made for unsupported platforms (pre OpenCL 1.2)
Update: I think there is a bug in cl2.hpp which despite setting minimum OpenCL version to 1.1 it still tries to use clRetain* on pre OpenCL1.2 devices when creating a command queue.
Setting minimum OpenCL version to 110 and version filtering works fine for me.
Complete working example:
#include "stdafx.h"
#include <vector>
#define CL_HPP_ENABLE_EXCEPTIONS
#define CL_HPP_TARGET_OPENCL_VERSION 200
#define CL_HPP_MINIMUM_OPENCL_VERSION 110
#include <CL/cl2.hpp>
using namespace std;
class TestClass
{
private:
std::vector<cl::CommandQueue> queues;
TestClass();
public:
static const TestClass& getInstance()
{
static TestClass instance;
return instance;
}
};
TestClass::TestClass()
{
std::vector<cl::Device> devices;
vector<cl::Platform> platforms;
cl::Platform::get(&platforms);
size_t x = 0;
for (; x < platforms.size(); ++x)
{
cl::Platform &p = platforms[x];
int v = cl::detail::getPlatformVersion(p());
short version_major = v >> 16;
if (version_major >= 2) // OpenCL 2.x
break;
}
if (x == platforms.size())
return; // no OpenCL 2.0 platform available
platforms[x].getDevices(CL_DEVICE_TYPE_GPU, &devices);
cl::Context context(devices);
cl::CommandQueue queue(context, devices[0]);
queues.push_back(queue);
}
int main()
{
TestClass::getInstance();
return 0;
}
Update2:
So what could be the reason for this and what are possible solutions?
I suspect it has something to do with the fact that my testclass is
static, but since it works with the executable I am at a loss what is
causing it.
TestClass static seems to be a reason. Looks like releasing memory is happening in wrong order when run from python. To fix that you may want to add a method which will have to be explicitly called to release opencl objects before python starts releasing memory.
static TestClass& getInstance() // <- const removed
{
static TestClass instance;
return instance;
}
void release()
{
queues.clear();
}
BOOST_PYTHON_MODULE(FrameWork)
{
TestClass::getInstance();
TestClass::getInstance().release();
}
"I would appreciate an answer that explains to me what the problem actually is and if there are ways to fix it."
First, let me say that doqtor already answered how to fix the issue -- by ensuring a well-defined destruction time of all used OpenCL resources. IMO, this is not a "hack", but the right thing to do. Trying to rely on static init/cleanup magic to do the right thing -- and watching it fail to do so -- is the real hack!
Second, some thoughts about the issue: the actual problem is even more complex than the common static initialization order fiasco stories. It involves DLL loading/unloading order, both in connection with python loading your custom dll at runtime and (more importantly) with OpenCL's installable client driver (ICD) model.
What DLLs are involved when running an application/dll that uses OpenCL? To the application, the only relevant DLL is the opencl.dll you link against. It is loaded into process memory during application startup time (or when your custom DLL which needs opencl is dynamically loaded in python).
Then, at the time when you first call clGetPlatformInfo() or similar in your code, the ICD logic kicks in: opencl.dll will look for installed drivers (in windows, those are mentioned somewhere in the registry) and dynamically load their respective dlls (using sth like the LoadLibrary() system call). That may be e.g. nvopencl.dll for nvidia, or some other dll for the intel driver you have installed. Now, in contrast to the relatively simple opencl.dll, this ICD dll can and will have a multitude of dependencies on its own -- probably using Intel IPP, or TBB, or whatever. So by now, things have become real messy already.
Now, during shutdown, the windows loader must decide which dlls to unload in which order. When you compile your example in a single executable, the number and order of dlls being loaded/unloaded will certainly be different than in the "python loads your custom dll at runtime" scenario. And that could well be the reason why you experience the problem only in the latter case, and only if you still have an opencl-context+commandqueue alive during shutdown of your custom dll. The destruction of your queue (triggered via clRelease... during static destruction of your testclass instance) is delegated to the intel-icd-dll, so this dll must still be fully functional at that time. If, for some reason, that is not the case (perhaps because the loader chose to unload it or one of the dlls it needs), you crash.
That line of thought reminded me of this article:
https://blogs.msdn.microsoft.com/larryosterman/2004/06/10/dll_process_detach-is-the-last-thing-my-dlls-going-to-see-right/
There's a paragraph, talking about "COM objects", which might be equally applicable to "OpenCL resources":
"So consider the case where you have a DLL that instantiates a COM object at some point during its lifetime. If that DLL keeps a reference to the COM object in a global variable, and doesn’t release the COM object until the DLL_PROCESS_DETACH, then the DLL that implements the COM object will be kept in memory during the lifetime of the COM object. Effectively the DLL implementing the COM object has become dependant on the DLL that holds the reference to the COM object. But the loader has no way of knowing about this dependency. All it knows is that the DLL’s are loaded into memory."
Now, I wrote a lot of words without coming to a definitive proof of what's actually going wrong. The main lesson I learned from bugs like these is: don't enter that snake pit, and do your resource-cleanup in a well-defined place like doqtor suggested. Good night.
I am facing a problem similar to the Py_initialize / Py_Finalize not working twice with numpy .. The basic coding in C:
Py_Initialize();
import_array();
//Call a python function which imports numpy as a module
//Py_Finalize()
The program is in a loop and it gives a seg fault if the python code has numpy as one of the imported module. If I remove numpy, it works fine.
As a temporary work around I tried not to use Py_Finalize(), but that is causing huge memory leaks [ observed as the memory usage from TOP keeps on increasing ]. And I tried but did not understand the suggestion in that link I posted. Can someone please suggest the best way to finalize the call while having imports such as numpy.
Thanks
santhosh.
I recently faced a very similar issue and developed a workaround that works for my purposes, so I thought I would write it here in the hope it might help others.
The problem
I work with some postprocessing pipeline for which I can write a own functor to work on some data passing through the pipeline and I wanted to be able to use Python scripts for some of the operations.
The problem is that the only thing I can control is the functor itself, which gets instantiated and destroyed at times beyond my control. I furthermore have the problem that even if I do not call Py_Finalize the pipeline sometimes crashes once I pass another dataset through the pipeline.
The solution in a Nutshell
For those who don't want to read the whole story and get straight to the point, here's the gist of my solution:
The main idea behind my workaround is not to link against the Python library, but instead load it dynamically using dlopen and then get all the addresses of the required Python functions using dlsym. Once that's done, one can call Py_Initialize() followed by whatever you want to do with Python functions followed by a call to Py_Finalize() once you're done. Then, one can simply unload the Python library. The next time you need to use Python functions, simply repeat the steps above and Bob's your uncle.
However, if you are importing NumPy at any point between Py_Initialize and Py_Finalize, you will also need to look for all the currently loaded libraries in your program and manually unload those using dlclose.
Detailed workaround
Loading instead of linking Python
The main idea as I mentioned above is not to link against the Python library. Instead, what we will do is load the Python library dynamically using dlopen():
#include
...
void* pHandle = dlopen("/path/to/library/libpython2.7.so", RTLD_NOW | RTLD_GLOBAL);
The code above loads the Python shared library and returns a handle to it (the return type is an obscure pointer type, thus the void*). The second argument (RTLD_NOW | RTLD_GLOBAL) is there to make sure that the symbols are properly imported into the current application's scope.
Once we have a pointer to the handle of the loaded library, we can search that library for the functions it exports using the dlsym function:
#include <dlfcn.h>
...
// Typedef named 'void_func_t' which holds a pointer to a function with
// no arguments with no return type
typedef void (*void_func_t)(void);
void_func_t MyPy_Initialize = dlsym(pHandle, "Py_Initialize");
The dlsym function takes two parameters: a pointer to the handle of the library that we obtained previously and the name of the function we are looking for (in this case, Py_Initialize). Once we have the address of the function we want, we can create a function pointer and initialize it to that address. To actually call the Py_Initialize function, one would then simply write:
MyPy_Initialize();
For all the other functions provided by the Python C-API, one can just add calls to dlsym and initialize function pointers to its return value and then use those function pointers instead of the Python functions. One simply has to know the parameter and return value of the Python function in order to create the correct type of function pointer.
Once we are finished with the Python functions and call Py_Finalize using a procedure similar to the one for Py_Initialize one can unload the Python dynamic library in the following way:
dlclose(pHandle);
pHandle = NULL;
Manually unloading NumPy libraries
Unfortunately, this does not solve the segmentation fault problems that occur when importing NumPy. The problems comes from the fact that NumPy also loads some libraries using dlopen (or something equivalent) and those do not get unloaded them when you call Py_Finalize. Indeed, if you list all the loaded libraries within your program, you will notice that after closing the Python environment with Py_Finalize, followed by a call to dlclose, some NumPy libraries will remain loaded in memory.
The second part of the solution requires to list all the Python libraries that remain in memory after the call dlclose(pHandle);. Then, for each of those libraries, grab a handle to them and then call dlcloseon them. After that, they should get unloaded automatically by the operating system.
Fortunately, there are functions under both Windows and Linux (sorry MacOS, couldn't find anything that would work in your case...):
- Linux: dl_iterate_phdr
- Windows: EnumProcessModules in conjunction with OpenProcess and GetModuleFileNameEx
Linux
This is rather straight forward once you read the documentation about dl_iterate_phdr:
#include <link.h>
#include <string>
#include <vector>
// global variables are evil!!! but this is just for demonstration purposes...
std::vector<std::string> loaded_libraries;
// callback function that gets called for every loaded libraries that
// dl_iterate_phdr finds
int dl_list_callback(struct dl_phdr_info *info, size_t, void *)
{
loaded_libraries.push_back(info->dlpi_name);
return 0;
}
int main()
{
...
loaded_libraries.clear();
dl_iterate_phdr(dl_list_callback, NULL);
// loaded_libraries now contains a list of all dynamic libraries loaded
// in your program
....
}
Basically, the function dl_iterate_phdr cycles through all the loaded libraries (in the reverse order they were loaded) until either the callback returns something other than 0 or it reaches the end of the list. To save the list, the callback simply adds each element to a global std::vector (one should obviously avoid global variables and use a class for example).
Windows
Under Windows, things get a little more complicated, but still manageable:
#include <windows.h>
#include <psapi.h>
std::vector<std::string> list_loaded_libraries()
{
std::vector<std::string> m_asDllList;
HANDLE hProcess(OpenProcess(PROCESS_QUERY_INFORMATION
| PROCESS_VM_READ,
FALSE, GetCurrentProcessId()));
if (hProcess) {
HMODULE hMods[1024];
DWORD cbNeeded;
if (EnumProcessModules(hProcess, hMods, sizeof(hMods), &cbNeeded)) {
const DWORD SIZE(cbNeeded / sizeof(HMODULE));
for (DWORD i(0); i < SIZE; ++i) {
TCHAR szModName[MAX_PATH];
// Get the full path to the module file.
if (GetModuleFileNameEx(hProcess,
hMods[i],
szModName,
sizeof(szModName) / sizeof(TCHAR))) {
#ifdef UNICODE
std::wstring wStr(szModName);
std::string tModuleName(wStr.begin(), wStr.end());
#else
std::string tModuleName(szModName);
#endif /* UNICODE */
if (tModuleName.substr(tModuleName.size()-3) == "dll") {
m_asDllList.push_back(tModuleName);
}
}
}
}
CloseHandle(hProcess);
}
return m_asDllList;
}
The code in this case is slightly longer than for the Linux case, but the main idea is the same: list all the loaded libraries and save them into a std::vector. Don't forget to also link your program to the Psapi.lib!
Manual unloading
Now that we can list all the loaded libraries, all you need to do is find among those the ones that come from loading NumPy, grab a handle to them and then call dlclose on that handle. The code below will work on both Windows and Linux, provided that you use the dlfcn-win32 library.
#ifdef WIN32
# include <windows.h>
# include <psapi.h>
# include "dlfcn_win32.h"
#else
# include <dlfcn.h>
# include <link.h> // for dl_iterate_phdr
#endif /* WIN32 */
#include <string>
#include <vector>
// Function that list all loaded libraries (not implemented here)
std::vector<std::string> list_loaded_libraries();
int main()
{
// do some preprocessing stuff...
// store the list of loaded libraries now
// any libraries that get added to the list from now on must be Python
// libraries
std::vector<std::string> loaded_libraries(list_loaded_libraries());
std::size_t start_idx(loaded_libraries.size());
void* pHandle = dlopen("/path/to/library/libpython2.7.so", RTLD_NOW | RTLD_GLOBAL);
// Not implemented here: get the addresses of the Python function you need
MyPy_Initialize(); // Needs to be defined somewhere above!
MyPyRun_SimpleString("import numpy"); // Needs to be defined somewhere above!
// ...
MyPyFinalize(); // Needs to be defined somewhere above!
// Now list the loaded libraries again and start manually unloading them
// starting from the end
loaded_libraries = list_loaded_libraries();
// NB: this below assumes that start_idx != 0, which should always hold true
for(std::size_t i(loaded_libraries.size()-1) ; i >= start_idx ; --i) {
void* pHandle = dlopen(loaded_libraries[i].c_str(),
#ifdef WIN32
RTLD_NOW // no support for RTLD_NOLOAD
#else
RTLD_NOW|RTLD_NOLOAD
#endif /* WIN32 */
);
if (pHandle) {
const unsigned int Nmax(50); // Avoid getting stuck in an infinite loop
for (unsigned int j(0) ; j < Nmax && !dlclose(pHandle) ; ++j);
}
}
}
Final words
The examples shown here capture the basic ideas behind my solution, but can certainly be improved to avoid global variables and facilitate ease of use (for example, I wrote a singleton class that handles the automatic initialization of all the function pointers after loading the Python library).
I hope this can be useful to someone in the future.
References
dl_iterate_phdr: https://linux.die.net/man/3/dl_iterate_phdr
PsAPI library: https://msdn.microsoft.com/en-us/library/windows/desktop/ms684894(v=vs.85).aspx
OpenProcess: https://msdn.microsoft.com/en-us/library/windows/desktop/ms684320(v=vs.85).aspx
EnumProcess: https://msdn.microsoft.com/en-us/library/windows/desktop/ms682629(v=vs.85).aspx
GetModuleFileNameEx: https://msdn.microsoft.com/en-us/library/windows/desktop/ms683198(v=vs.85).aspx
dlfcn-win32 library: library: https://github.com/dlfcn-win32/dlfcn-win32
I'm not quite sure how you don't seem to understand the solution posted in Py_initialize / Py_Finalize not working twice with numpy. The solution posted is quite simple: call Py_Initialize and Py_Finalize only once for each time your program executes. Do not call them every time you run the loop.
I assume that your program, when it starts, runs some initialization commands (which are only run once). Call Py_Initialize there. Never call it again. Also, I assume that when your program terminates, it has some code to tear down things, dump log files, etc. Call Py_Finalize there. Py_Initialize and Py_Finalize are not intended to help you manage memory in the Python interpreter. Do not use them for that, as they cause your program to crash. Instead, use Python's own functions to get rid of objects you don't want to keep.
If you really MUST create a new environment every time you run your code, you can use Py_NewInterpreter and to create a sub-interpreter and Py_EndInterpreter to destroy that sub-interpreter later. They're documented near the bottom of the Python C API page. This works similarly to having a new interpreter, except that modules are not re-initialized each time a sub-interpreter starts.