aarch64: libgomp.so.1: cannot allocate memory in static TLS block

[ric96]

Problem is caused by this optional flag: -DWITH_OPENMP=ON
Comment with workaround: #14884 (comment)
Comment with investigation: #14884 (comment)

---

Original description

I am compiling opencv from source in order to enable multi-core support on a aarch64 platform.

I am building open cv with the following config:

cmake -D CMAKE_BUILD_TYPE=RELEASE -DBUILD_SHARED_LIBS=OFF -DBUILD_EXAMPLES=OFF -DBUILD_opencv_apps=OFF -DBUILD_DOCS=OFF -DBUILD_PERF_TESTS=OFF -DBUILD_TESTS=OFF -DCMAKE_INSTALL_PREFIX=/usr/local -DENABLE_PRECOMPILED_HEADERS=OFF -DWITH_LIBV4L=ON -DWITH_QT=ON -DWITH_OPENGL=ON -DFORCE_VTK=ON -DWITH_TBB=ON -DWITH_GDAL=ON -DWITH_XINE=ON -DWITH_OPENMP=ON -DWITH_GSTREAMER=ON -DWITH_OPENCL=ON -DOPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules ../

but every time I import it using python, I get the following error. A lot of the results on-line seems to be very vague and/or point towards a bug in gcc 4.2 fixed in 4.3 and I am currently on 8.x provided by debian repo and also 9.0.1 built from upstream source:

root@linaro-developer:~# python3
Python 3.7.3 (default, Apr  3 2019, 05:39:12) 
[GCC 8.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import cv2
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.7/dist-packages/cv2/__init__.py", line 89, in <module>
    bootstrap()
  File "/usr/local/lib/python3.7/dist-packages/cv2/__init__.py", line 79, in bootstrap
    import cv2
ImportError: /usr/lib/aarch64-linux-gnu/libgomp.so.1: cannot allocate memory in static TLS block

This is on opencv master branch as well as 4.0.1

[ric96]

I can now confirm that OpenCV 3.2 built with openmp is working fine on the exact same platform.

[mshabunin]

I just wanted to point that it is not necessary to enable TBB or OpenMP in order to enable multicore support. Default pthreads backend can be good enough for most situations.

[jmliu1983]

I have the same issue.
Hardware: Jetson Xavier
IDE: Spyder 3.2.6
Python 3.6.8
It's strange that this happened only in Spyder 3.2.6, when I import cv2 in consol, it works well.

[twmr]

See pytorch/pytorch#2575

[ivanfarkas2]

@ric96 Sahaj, did @Thisch suggestion worked for you?

[matallan]

I had a similar issue where I could import cv2 in consol but not in a script. I was importing tensorflow before cv2. For some reason switching the order (importing cv2 first) worked. Not sure why - but if it helps someone else 👍

[AhmedHisham1]

I had a similar issue where I could import cv2 in consol but not in a script. I was importing tensorflow before cv2. For some reason switching the order (importing cv2 first) worked. Not sure why - but if it helps someone else 👍

@matallan

I had the same problem on Nvidia Jetson Nano device.
Switching the import order has worked for me too.
Seems to be a problem with the embedded device itself. Perhaps because I'm not running it on the 10W power mode.

[tlhcelik]

I had a similar issue where I could import cv2 in consol but not in a script. I was importing tensorflow before cv2. For some reason switching the order (importing cv2 first) worked. Not sure why - but if it helps someone else 👍

it's working to me. Thanks

[Kroj-0]

I had a similar issue where I could import cv2 in consol but not in a script. I was importing tensorflow before cv2. For some reason switching the order (importing cv2 first) worked. Not sure why - but if it helps someone else 👍

no idea why, fixed it for me too

[serenadeRay]

keras-team/keras-tuner#317

[PetrDvoracek]

have you tried export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libgomp.so.1 ?

[emrectn]

I had a similar issue where I could import cv2 in consol but not in a script. I was importing tensorflow before cv2. For some reason switching the order (importing cv2 first) worked. Not sure why - but if it helps someone else

It is working to me. Thank you

[Qengineering]

Most likely cause here

That is a bug in whatever dlopened libgomp.so.1 and before doing that ate all the preallocated TLS area.
The GNU TLS2 model which I'm afraid aarch64 uses unfortunately eats from the same TLS preallocated pool as libraries that require static TLS like libgomp, where it is performance critical to have it as static TLS.
Either don't dlopen libgomp, or LD_PRELOAD it, link it with the application that dlopens it, cut down the uses of other TLS or dlopen libgomp earlier.
There is nothing that can be done on the gcc side.

Three possible solutions:

1. Disable the use of OpenMP during installation of OpenCV. Place the arguments -DBUILD_OPENMP=OFF -DWITH_OPENMP=OFF in your CMAKE command line. OpenCV will now use pthread. Performance losses are hardly noticeable. (Of course, only possible if you install OpenCV from scratch and not by pip or sudo apt install)

2. Use TBB as parallel framework. sudo apt-get install libtbb2 libtbb-dev loads the needed libraries.
   -DBUILD_TBB=ON -DWITH_TBB=ON in your CMAKE command. TBB overrules OpenMP.
   (Again, only possible if you install OpenCV from scratch)

3. As stated above, make sure another application uses libgomp before OpenCV. This can be the import of TensorFlow, or the LD_PRELOAD construction.