Bug 1179055

+++ This bug was initially created as a clone of Bug #1172834 +++

Description of problem:
ATLAS xerbla exits on error. Numpy normally overrides that function, but that seems to no longer work in fedora 21.

How reproducible: Always


Steps to Reproduce:

charris@localhost [~]$ python
Python 2.7.8 (default, Nov 10 2014, 08:19:18) 
[GCC 4.9.2 20141101 (Red Hat 4.9.2-1)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy as np
>>> np.linalg.lstsq([[np.nan]]*2, [1]*2)
 ** On entry to DLASCL parameter number  4 had an illegal value
charris@localhost [~]$ 

Expected results:
Raises an error and Python continues.


Additional info:
You might be building ATLAS with -Wl,-Bsymbolic-functions, which will cause the problem.

--- Additional comment from charles harris on 2014-12-10 16:15:15 EST ---

The numpy is locally built and linked to libtatlas. I note that Fedora numpy is linked to libsatlas, not sure why.

--- Additional comment from Frantisek Kluknavsky on 2014-12-12 08:31:08 EST ---

I can confirm the different behavior of numpy between f20 and f21. Problem does not seem to be in atlas, my mini-example with custom xerbla works fine, with both satlas and tatlas. Compiled with Makefile:
CC=gcc
F90=gfortran
test: main.o routines.o Makefile
        $(CC) -o test main.o routines.o -lm -lgfortran -L/usr/lib64/atlas -ltatlas
%.o: %.c
        $(CC) -Wall -O2 -g -c $< -o $@
%.o: %.f90
        $(F90) -Wall -O2 -g -c $< -o $@

Call dgemm('','n',m,n,k,1.0,a,m,b,k,0.0,c,m) with wrong first parameter kills the process - default xerbla. With
 subroutine xerbla(SRNAME, INFO)
  CHARACTER*(*)      SRNAME
  INTEGER            INFO
  PRINT *, "Hello World!"
 end subroutine xerbla
in routines.f90 it prints the hello message and process continues - successful override.

Maybe something changed in numpy?

I do not know why numpy is linked to serial and not threaded atlas in Fedora. One probable answer might be that threaded atlas is currently broken on architectures with weak cache coherency, or maybe because it can not detect the number of cores at runtime and to get real benefits you need to recompile atlas yourself, or maybe because someone wants to use numpy as a generic library and do his own parallelization. Many problems with threaded atlas for little benefit.

--- Additional comment from Orion Poplawski on 2014-12-12 22:36:07 EST ---

This because multiple numpy shared libraries are linked against atlas:

rpm -ql numpy | grep -F .so | while read so; do ldd $so | grep atlas && echo $so; done
        libsatlas.so.3 => /usr/lib64/atlas/libsatlas.so.3 (0x00007fd169763000)
/usr/lib64/python2.7/site-packages/numpy/core/_dotblas.so
        libsatlas.so.3 => /usr/lib64/atlas/libsatlas.so.3 (0x00007f7135291000)
/usr/lib64/python2.7/site-packages/numpy/linalg/_umath_linalg.so
        libsatlas.so.3 => /usr/lib64/atlas/libsatlas.so.3 (0x00007f67f2df9000)
/usr/lib64/python2.7/site-packages/numpy/linalg/lapack_lite.so

But the xerbla_ override is not in _dotblas.so and that is getting loading first.

--- Additional comment from Orion Poplawski on 2014-12-12 22:48:27 EST ---

I've updated the upstream issue tracker:
https://github.com/numpy/numpy/issues/5362

Would have been nice to have mentioned that link in the original report.

--- Additional comment from charles harris on 2014-12-19 20:13:23 EST ---

PR https://github.com/numpy/numpy/pull/5382 for numpy fixes this, but you will probably need to fix the fedora version of numpy yourself. All that seems to be needed is to move the linalg import in numpy/__init__.py before the import of core. The fix *does not work* if site.cfg is used for the setup instead of being automatically detected as by PR https://github.com/numpy/numpy/pull/5364.

--- Additional comment from charles harris on 2014-12-19 20:22:52 EST ---

NVM, still broken, needs work.

--- Additional comment from charles harris on 2014-12-24 14:01:40 EST ---

See https://github.com/numpy/numpy/pull/5392. That patch is for 1.9.x, but the same fix is needed for the 1.8.2 that is in Fedora.

--- Additional comment from marcindulak on 2015-01-02 09:23:06 EST ---

(In reply to charles harris from comment #7)
> See https://github.com/numpy/numpy/pull/5392. That patch is for 1.9.x, but
> the same fix is needed for the 1.8.2 that is in Fedora.

the problem is present also on RHEL7: numpy-1.7.1-10.el7.x86_64