The solver hangs if I compile with threads and run it from python.

[stschubert]

Is there anything new about the below reported issue? I'm asking since I have the same problem.

Update: Looks like the solver hangs if I compile with threads and run it from python. I removed the cmake flags above.

Originally posted by @NikolausDemmel in #7 (comment)

[stschubert]

By now, I could figure out that there is a problem in minisam/python/minisam_wrapper/factor.cpp in PYBIND11_OVERLOAD_PURE(Eigen::VectorXd, Factor, error, variables); (line 70). However, I don't know how pybind works...

[dongjing3309]

Hi @stschubert , my feeling is I must have a bug in multi-threading and causes deadlock. Do you have the example here which will helps me debug a lot? Thanks!

[NikolausDemmel]

I just came across this discussion about python bindings and multi-threading, and thought that it might be related.

[Edwinem]

I was able to reproduce the issue with the gps_factor_example.py. I can also confirm that it has to do with the GIL.

Running the code through a debugger it deadlocks currently on

minisam/minisam/core/FactorGraph.cpp

Line 83 in eda47f9

linthreads[i].join();

and

https://github.com/pybind/pybind11/blob/5b4751af269afab9a2e40a92b486bd29214f352b/include/pybind11/pybind11.h#L1911

It can be fixed by releasing the GIL before the thread joins. As seen below.

  pybind11::gil_scoped_release release;
  // wait threads to finish
  for (int i = 0; i < MINISAM_WITH_MULTI_THREADS_NUM; i++) {
    linthreads[i].join();
  }

This has to be done in both FactorGraph.cpp and linearization.cpp.

Note this now creates a dependency on Python and Pybind11 in minisam. So you probably have to add some new compile time flags where if the python package is enabled then this line is added.

Update:

A better fix is provided in #10. It is cleaner and does not require a Python dependency.

[dongjing3309]

should be fixed by PR1