Issue with OpenMP enabled python binary wheels

[pramodk]

Describe the issue

Like OpenACC/CUDA, ABI compatibility of OpenMP and shipping duplicate OpenMP runtime library is problematic.

To Reproduce

• Build the wheel on manylinux2014_aarch64
• Try to use the wheel on Ubuntu20 Aarch64 (circle-ci) gives:

(nrn_test_venv_39) circleci@ip-172-28-90-45:~/project$ python test/coreneuron/test_direct.py
NEURON: unable to open font "*helvetica-medium-r-normal*--14*", using "fixed"
/lib/aarch64-linux-gnu/libgomp.so.1: cannot allocate memory in static TLS block
NEURON: Could not dlopen CoreNEURON mechanism library :  /home/circleci/project/aarch64/libcorenrnmech.so
 near line 0
 create soma
            ^
        ParallelContext[0].psolve(5)
Traceback (most recent call last):
  File "/home/circleci/project/test/coreneuron/test_direct.py", line 92, in <module>
    test_direct_memory_transfer()
  File "/home/circleci/project/test/coreneuron/test_direct.py", line 70, in test_direct_memory_transfer
    run(mode)
  File "/home/circleci/project/test/coreneuron/test_direct.py", line 52, in run
    pc.psolve(h.tstop)
RuntimeError: hocobj_call error

Expected behavior

Wheel should work without any issue.

Additional context

• See relevant discussion in aarch64: libgomp.so.1: cannot allocate memory in static TLS block opencv/opencv#14884 and OpenMP related
• auditwheel packages extra libraries like OpenMP

$ ldd aarch64/libcorenrnmech.so
	linux-vdso.so.1 (0x0000ffffb379e000)
	libcoreneuron.so => /home/circleci/project/nrn_test_venv_39/lib/python3.9/site-packages/neuron/.data/lib/libcoreneuron.so (0x0000ffffb35ce000)
	libstdc++.so.6 => /lib/aarch64-linux-gnu/libstdc++.so.6 (0x0000ffffb33dc000)
	libm.so.6 => /lib/aarch64-linux-gnu/atomics/libm.so.6 (0x0000ffffb332e000)
	libgomp.so.1 => /lib/aarch64-linux-gnu/libgomp.so.1 (0x0000ffffb32e0000)
	libc.so.6 => /lib/aarch64-linux-gnu/atomics/libc.so.6 (0x0000ffffb316f000)
	/lib/ld-linux-aarch64.so.1 (0x0000ffffb376e000)
	libdl.so.2 => /lib/aarch64-linux-gnu/atomics/libdl.so.2 (0x0000ffffb315b000)
	libgomp-d22c30c5.so.1.0.0 => /home/circleci/project/nrn_test_venv_39/lib/python3.9/site-packages/neuron/.data/lib/../../../NEURON_nightly.libs/libgomp-d22c30c5.so.1.0.0 (0x0000ffffb3107000)
	libgcc_s.so.1 => /lib/aarch64-linux-gnu/libgcc_s.so.1 (0x0000ffffb30e2000)
	libpthread.so.0 => /lib/aarch64-linux-gnu/atomics/libpthread.so.0 (0x0000ffffb30b3000)

On the host machine OpenMP runtime exists in:

(nrn_test_venv_39) circleci@ip-172-28-90-45:~/project$ ls /usr/lib/aarch64-linux-gnu/libgomp*
/usr/lib/aarch64-linux-gnu/libgomp.so.1  /usr/lib/aarch64-linux-gnu/libgomp.so.1.0.0

And forcing this library makes thing "better":

(nrn_test_venv_39) circleci@ip-172-28-90-45:~/project$ LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libgomp.so.1 python test/coreneuron/test_direct.py
NEURON: unable to open font "*helvetica-medium-r-normal*--14*", using "fixed"

================ MECHANISMS COUNT BY TYPE ==================
  Id                 Name      Count
   0               (null)          0
   1               (null)          0
   2           morphology          0
   3          capacitance          1
   4                  pas          0
   5        extracellular          0
   6              fastpas          0
   7               IClamp          1
   8         AlphaSynapse          0
   9               ExpSyn          0
  10              Exp2Syn          0
  11              SEClamp          0
  12               VClamp          0
  13               OClamp          0
  14              APCount          0
  15               na_ion          1
  16                k_ion          1
  17                   hh          1
  18              NetStim          0
  19             IntFire1          0
  20             IntFire2          0
  21             IntFire4          0
  22     PointProcessMark          0
  23          PatternStim          0
  24                  ba1          0
  25                  ba0          0
  26        ForNetConTest          0
  27         IntervalFire          0
  28                DAsyn          0
  29               Sample          1
  30               Bounce          0
=============================================================
Memory size information for all NrnThreads per rank
------------------------------------------------------------------
                 field          min          max          avg
                n_cell            1            1            1.00
....

@ferdonline : we have similar issue with GPU (OpenACC runtime) and you can see @olupton's hack in https://github.com/neuronsimulator/nrn/pull/1661/files. One of previous discussion was if we should workaround this in auditwheel.

[alexsavulescu]

For reference: pypa/auditwheel#368