helping MUMPS configure find LAPACK

[rdzman]

I'm trying to add an IPOPT MEX interface to a GNU Octave Docker image. Using apt-get coinor-libipopt-dev resulted in a MEX file that causes a segfault. I'm guessing it may be because of mismatched BLAS, LAPACK, etc. (though this is pure conjecture ... not sure how to confirm).

So, I'm now trying to build IPOPT from scratch using coinbrew and running into an issue with the configure for MUMPS not being able to find LAPACK. I'm assuming it's already been used to build Octave and is around somewhere, but I'm not sure where. According to this docker file, they are installing OpenBLAS and SuiteSparse in /usr. And /usr/lib does contain libopenblas.so, for example, but I'm not familiar with which library specifically would contain LAPACK.

I've tried setting the with_lapack_lflags env variable to several different values with no luck. I always get an error like, e.g. ...

configure: error: Cannot link to user-specified Lapack  -L/usr/lib -llibopenblas.so.

Any suggestions for what with_lapack_lflags should be?

[rdzman]

Ok, I figured that out. Setting with_lapack_lflags to -lopenblas allows it to configure and build everything just fine now. However, it's segfaulting on the unit tests now.

Any suggestions on how to track that down?

##################################################
### Building Ipopt 3.14.4 
##################################################


##################################################
### Running Ipopt unit test 
##################################################

./run_unitTests: line 21: 18957 Segmentation fault      $@ > tmpfile 2>&1
./run_unitTests: line 21: 18972 Segmentation fault      $@ > tmpfile 2>&1
./run_unitTests: line 21: 18987 Segmentation fault      $@ > tmpfile 2>&1
./run_unitTests: line 21: 19002 Segmentation fault      $@ > tmpfile 2>&1
./run_unitTests: line 21: 19017 Aborted                 $@ > tmpfile 2>&1
./run_unitTests: line 21: 19035 Segmentation fault      $@ > tmpfile 2>&1
./run_unitTests: line 21: 19050 Segmentation fault      $@ > tmpfile 2>&1
./run_unitTests: line 21: 19065 Segmentation fault      $@ > tmpfile 2>&1
./run_unitTests: line 21: 19080 Segmentation fault      $@ > tmpfile 2>&1
make[1]: *** [Makefile:816: test] Error 1
make: *** [Makefile:848: test] Error 2

[rdzman]

A bit more info ... it looks like the verbose output of the Java based test has the most info ...

# A fatal error has been detected by the Java Runtime Environment:
#
#  SIGSEGV (0xb) at pc=0x00007fa5f7089d87, pid=22130, tid=22131
#
# JRE version: OpenJDK Runtime Environment (11.0.11+9) (build 11.0.11+9-Ubuntu-0ubuntu2.20.04)
# Java VM: OpenJDK 64-Bit Server VM (11.0.11+9-Ubuntu-0ubuntu2.20.04, mixed mode, sharing, tiered, compressed oops, g1 gc, linux-amd64)
# Problematic frame:
# C  [libopenblas.so.0+0x179dd87]  dcopy_k_HASWELL+0x387

And another test says ...

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

Backtrace for this error:
#0  0x7f04294c2d5a
#1  0x7f04294c1ef5
#2  0x7f04292f620f
#3  0x7f04280cbd87

[svigerske]

It looks like they build OpenBlas with an interface for 64-bit integers for Octave:

https://github.com/gnu-octave/docker/blob/dd3d3b6f4b9202423275424ecb77c7b3bb57f7ed/build-octave-6.docker#L108-L122

But Ipopt and MUMPS assume 32-bit integers (or actually int) by default.
That's the same problem users sometimes have with MATLAB/Ipopt interfaces.

If you build MUMPS and Ipopt against a Blas/Lapack with int for integers instead, you should use static libraries for Blas/Lapack to avoid that the Blas/Lapack routines from Octave are still used (when they are loaded first).
Or you could build Ipopt and MUMPS to use 64-bit integers, too (https://coin-or.github.io/Ipopt/INSTALL.html#INT64_BUILD). But this comes with increased memory usage as every integer in Ipopt will be 64-bit then.

[siko1056]

Here the author of the Octave Docker images ✋ I just finished answering the question of @rdzman :

gnu-octave/docker#5 (comment)

@svigerske your analysis of the problem is correct and the solution is provided above. You have a great Ipopt documentation regarding ilp64 compilation 👍 Sorry for the duplicated effort looking into this (@rdzman please avoid crossposing, in FLOSS one cannot always receive a reply within "one unpaid business day" 😇 ).

[rdzman]

Thanks a ton to both of you, and sorry for the duplicate effort. Now I can build a working IPOPT, but the MEX build is failing. I'll post all followups in gnu-octave/docker#5.