Numpy: import error from numpy

Created on 24 Dec 2016  路  4Comments  路  Source: numpy/numpy

Hi,
When I write:"from numpy import *" from code and execute the code,then meet the following error:

Traceback (most recent call last):

File "rgbtoyuv.py", line 2, in
import numpy as np
File "/usr/local/lib/python3.5/site-packages/numpy-1.11.2-py3.5-linux-x86_64.egg/numpy/__init__.py", line 163, in
from . import random
File "/usr/local/lib/python3.5/site-packages/numpy-1.11.2-py3.5-linux-x86_64.egg/numpy/random/__init__.py", line 99, in
from .mtrand import *
ImportError: /usr/local/lib/python3.5/site-packages/numpy-1.11.2-py3.5-linux-x86_64.egg/numpy/random/mtrand.cpython-35m-x86_64-linux-gnu.so: undefined symbol: PyFPE_jbuf

Platform ubuntu16.04 x86_64

python version: 3.5.2

numpy version:1.11.2 and try to version:1.9.0 also,but installed failed.

Please help me find the reason,thanks a lot.

Most helpful comment

Seems like a bug report to me?

Some clues as to what's going on here:

CPython has some optional code for catching SIGFPE and converting it into an exception. To use it, you wrap C-level floating point computations in PyFPE_START_PROTECT/PyFPE_END_PROTECT. If CPython was compiled with --with-fpectl (which corresponds to having a #define WANT_SIGFPE_HANDLER in pyconfig.h), then PyFPE_jbuf is exported as a global symbol, and these macros expand into some clever stuff that references this symbol. If CPython was compiled without --with-fpectl, then the symbol isn't exported and the macros become no-ops. So this means that CPython actually has a different, incompatible ABI depending on whether it was compiled with or without --with-fpectl. Joy.

(It looks like you can check which sort of CPython you have by doing import fpectl -- if this succeeds then you have a --with-fpectl CPython.)

Numpy itself doesn't use these macros. But Cython autogenerated code does use them.

Conclusion: if you compile a Cython module using a CPython that was built with --with-fpectl, then it will crash when you try to run it on a CPython that was built without --with-fpectl.

Quick checking of some common linux CPython builds to see which ones use --with-fpectl:

  • Manylinux: no
  • Current Debian testing: yes
  • Ubuntu 12.04: yes
  • Ubuntu 16.04: yes
  • Conda: no (checked 2.7 and 3.5)
  • Fedora 25: no

@kevinzhai80: Did you by some chance build numpy from source for some reason, using the Ubuntu-installed Python, and then try to import it from a conda-installed Python?

All 4 comments

Where did you get numpy? BTW, questions like this should be asked on the mailing list, github issues is for bugs

Seems like a bug report to me?

Some clues as to what's going on here:

CPython has some optional code for catching SIGFPE and converting it into an exception. To use it, you wrap C-level floating point computations in PyFPE_START_PROTECT/PyFPE_END_PROTECT. If CPython was compiled with --with-fpectl (which corresponds to having a #define WANT_SIGFPE_HANDLER in pyconfig.h), then PyFPE_jbuf is exported as a global symbol, and these macros expand into some clever stuff that references this symbol. If CPython was compiled without --with-fpectl, then the symbol isn't exported and the macros become no-ops. So this means that CPython actually has a different, incompatible ABI depending on whether it was compiled with or without --with-fpectl. Joy.

(It looks like you can check which sort of CPython you have by doing import fpectl -- if this succeeds then you have a --with-fpectl CPython.)

Numpy itself doesn't use these macros. But Cython autogenerated code does use them.

Conclusion: if you compile a Cython module using a CPython that was built with --with-fpectl, then it will crash when you try to run it on a CPython that was built without --with-fpectl.

Quick checking of some common linux CPython builds to see which ones use --with-fpectl:

  • Manylinux: no
  • Current Debian testing: yes
  • Ubuntu 12.04: yes
  • Ubuntu 16.04: yes
  • Conda: no (checked 2.7 and 3.5)
  • Fedora 25: no

@kevinzhai80: Did you by some chance build numpy from source for some reason, using the Ubuntu-installed Python, and then try to import it from a conda-installed Python?

Hi njsmith,
Thanks for your information,this issue is resolved.

Thanks njsmith. Also: Ubuntu 17.10 Artful no. Upgrading caused the error:
$ python
Python 2.7.14 (default, Sep 23 2017, 22:06:14)
[GCC 7.2.0] on linux2
Type "help", "copyright", "credits" or "license" for more information.

import fpectl
Traceback (most recent call last):
File "", line 1, in
ImportError: No module named fpectl

Was this page helpful?
0 / 5 - 0 ratings