Could not load shared library about eqcorrscan HOT 11 CLOSED

from eqcorrscan.

Ok I ran the build_ext command, then tried to run the tests and get the same error. This time I installed eqcorrcan with the develop flag, not sure if that makes a difference.

from eqcorrscan.

from eqcorrscan.

Ah, on reading more closely, it looks like you don't have the openmp libraries installed? I had thought this was standard in gcc, more info to follow (and will update docs once we get this working).

p.s. eqcorrscan is now on condo-forge, and comes with fftw and openmp when installing from there, but you can't run the tests from there...

from eqcorrscan.

It looks like a linking error to the openMP libraries, but I don't know why it would be (tests on 16.04 run fine with standard setups). Can I check that gcc is the compiler that gets called by setup.py build_ext? It would usually be gcc unless you built your Python in an interesting way (although the current method also works for intel compilers).

from eqcorrscan.

I am just running anaconda with python version 3.6, shouldn't be anything unusual about it.

It turns out eqcorrscan/lib/libutils.cpython-36m-x86_64-linux-gnu.so doesn't exist, not sure what that means as far as possible failures in the compile chain.

Edit:
eqcorrscan/lib/libutils.cpython-36m-x86_64-linux-gnu.so does exist, I was erroneously looking at a different instance of eqcorrscan. When running the tests in the correct directory I am having the same issue though, undefined symbol.

How should I go about testing if gcc is the compiler being used? I stepped through finalize_options function in setup.py with the debugger and compiler variable was set to "unix" if that is useful.

Edit2:

conda list | grep gcc

produces this:

libgcc 4.8.5 2

This might be a related issue but conda installing gcc did not fix the issue for me. I will keep digging.

from eqcorrscan.

from eqcorrscan.

Okay, so I can recreate this error when I install gcc using conda, rather than using the system version of gcc. What I actually did was:

1. create a new conda env (without installing gcc via conda)
2. build eqcorrscan using pip

• tests pass

4. installed gcc via conda

• tests fail with missing reference to GOMP functions

5. installed openMP from conda

• tests fail with same errors

7. clean out built files (remove build directory and any possible shared object in lib)
8. rebuild eqcorrscan with pip

• tests pass

As follows:

conda create -n eqcorrscan python=3.6 colorama numpy scipy matplotlib obspy bottleneck pyproj
source activate eqcorrscan
pip install .  #run from git repo directory
python eqcorrscan/tests/correlate_test.py
# Pass
conda install gcc
pip install .  -U --no-deps
python eqcorrscan/tests/correlate_test.py
# Fail
conda install openmp
pip install .  -U --no-deps
python eqcorrscan/tests/correlate_test.py
# Fail
rm -r build
rm eqcorrscan/lib/libutils.cpython-*.so
pip install .  -U --no-deps
python eqcorrscan/tests/correlate_test.py
# Pass

In my case, the library built fine, but was built with a different version of libgcc and libgomp than in the conda env, which is fine - I don't know if this holds for you? Did you install gcc via conda after building the libraries?

You can check which gcc is being used by running which gcc.

Also, if I make an environment including gcc:

conda create -n eqcorrscan python=3.6 colorama numpy scipy matplotlib obspy bottleneck pyproj gcc

It runs fine (and builds using the conda gcc).

from eqcorrscan.

Also note that you should be fine using the system gcc (which should have happened using the instructions on the website, so I don't know what happened there originally?).
When you update from the git repo you will need to clean out the built shared objects first, otherwise they will not be rebuilt.

from eqcorrscan.

Okay, it seems to work now with your suggestions. I think my original attempt to install eqcorrscan, without a virtual env, would have been similar to where you were at in step 4 listed above. The build directory would have existed, which was probably the reason the subsequent efforts failed with environments that should have worked had the build been removed.

Since compilation time is very small maybe build should be removed by default?

Thanks for your help.

from eqcorrscan.

I think removing build by default would be a good idea - might have to dig into setup-tools a little more to work out how that should be done.

from eqcorrscan.