MKL Pardiso error during symbolic factorization -3. (reordering_phase) error_num= -180 about osqp-python

[bstellato]

This is now fixed in the develop branch in 60f307b. It is related to the fact that numpy is compiled with 32bit ints support by default (see the links above).

Therefore, anaconda links numpy to the 32bit ints MKL library. If OSQP tries to use the same library with 64bit ints it creates memory issues. Too bad MKL does not detect that and reports a wrong error.

We now disabled 64bit integer support in python by default. It can be enabled with a simple flag in setup.py.

[rwolst]

I am having a similar issue when using the osqp from pypi. If I use conda then I end up getting a seg fault, explained here: osqp/osqp#65.

[bstellato]

The coredump should be fixed in the develop branch. However, there is an issue with using the anaconda mkl distribution. I have no problems on my mac with anaconda and numpy not linked to mkl

In [2]: numpy.show_config()
blas_mkl_info:
  NOT AVAILABLE
...
lapack_mkl_info:
  NOT AVAILABLE

However, when I try on another system where numpy.show_config() returns something different than NOT AVAILABLE I get the symbolic factorization error -3 as shown in osqp/osqp#65

In particular, if I enable the output from pardiso with msglvl = 1, I get the output

*** Error in PARDISO  (        reordering_phase) error_num= -180
*** error PARDISO: reordering, symbolic factorization

It seems related to some linking issue since the error appears also here. However, in the C library and in other interfaces I do not see it happening. I strongly suspect it is related to how anaconda manages the python/numpy mkl linkage.

[bstellato]

Possiibly related issues:

JuliaPy/PyCall.jl#443
ContinuumIO/anaconda-issues#3823
numpy/numpy#5906

[bstellato]

Other very related issue. Other people getting the same exact problem in Julia: JuliaSparse/Pardiso.jl#28