gsl build failure on Centos 6

[Nathan Dunfield]

I am trying to build Sage 9.1 from source on Scientific Linux 6.10 (=Centos 6) using gcc 7.5 provided by Conda, and am having problems getting gsl to build:

[gsl-2.5] /data/keeling/a/nmd/miniconda3/envs/sage_base_lite4/bin/../lib/gcc/x86_64-conda_cos6-linux-gnu/7.5.0/../../../../x86_64-conda_cos6-linux-gnu/bin/ld: /data/keeling/a/nmd/pkgs/sage-9.1/local/lib/libopenblas.so: undefined reference to `_gfortran_concat_string'  

Attached are the logs for gsl, openblas, and config.  The config command was:

./configure --with-system-gmp=no --with-system-openblas=no --with-system-gfan=no --with-system-iconv=no --with-system-libffi=no --with-system-sqlite=no

Thanks,

Nathan

P.S. In case you are wondering why I don't just use Sage via Conda, the answer is (a) some nodes on this cluster are old enough the Sage-on-conda binaries occasionally generate exceptions from missing instructions and (b) I need to install certain Sage optional packages, specifically Normaliz.

[Nathan Dunfield]

I am trying to build Sage 9.1 from source on Scientific Linux 6.10 (=Centos 6) using gcc 7.5 provided by Conda, and am having problems getting gsl to build:

[gsl-2.5] /data/keeling/a/nmd/miniconda3/envs/sage_base_lite4/bin/../lib/gcc/x86_64-conda_cos6-linux-gnu/7.5.0/../../../../x86_64-conda_cos6-linux-gnu/bin/ld: /data/keeling/a/nmd/pkgs/sage-9.1/local/lib/libopenblas.so: undefined reference to `_gfortran_concat_string'  

Attached are the logs for gsl, openblas, and config.  The config command was:

Sorry, I should also have mentioned that I modified the environment variables FFLAGS and DEBUG_FFLAGS to remove "-fopenmp" before running "make" in the "sage-9.1" directory as openmp was causing problems with other spkgs.

Nathan 

[Dima Pasechnik]

Can you use gsl provided by Conda?

If you run can Sage's ./configure with conda active then it will list "system" packages to install.

--
You received this message because you are subscribed to the Google Groups "sage-devel" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sage-devel+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/sage-devel/dccc298b-ad86-43af-8d60-74cff1c53f8co%40googlegroups.com.

[Isuru Fernando]

This is an issue with the openblas build. You might need https://github.com/conda-forge/openblas-feedstock/blob/master/recipe/0002-Fix-gfortran-detection-for-ctng-based-cross-compilers.patch  (That patch was not sent upstream because it breaks other gcc builds. Need to figure out a patch that works for all)

Also, I'm curious to hear more about the illegal instructions. We can fix that if you report them. I'm a conda-forge maintainer.

We also have Normaliz packaged in conda-forge.

Isuru

To view this discussion on the web visit https://groups.google.com/d/msgid/sage-devel/CAAWYfq3mKj5GwW8Z312yEuoZBoP6fwe0FAvsmfJZd-mjeg2C5Q%40mail.gmail.com.

[Nathan Dunfield]

On Friday, July 24, 2020 at 10:43:40 AM UTC-5, Dima Pasechnik wrote:

Can you use gsl provided by Conda?

If you run can Sage's ./configure with conda active then it will list "system" packages to install.

I don't think so because having conda install gsl will cause it to install conda's openblas.  Originally, I tried the maximal-conda approach, installing as many of Sage's dependencies that way as possible (following the sage conda feedstock script), but then giac wouldn't build.  I then slowly reduced the number of conda packages installed, but there were (sometimes different) build problems for Sage with everything I've tried.  Hence the "sage-as-distribution" approach I'm trying currently, which has worked for me in the past on this system, with many version of  Sage <= 8.4 (though I never could get Sage 8.9 to build).

Nathan

[Nathan Dunfield]

On Friday, July 24, 2020 at 10:58:12 AM UTC-5, Isuru Fernando wrote:

This is an issue with the openblas build. You might need https://github.com/conda-forge/openblas-feedstock/blob/master/recipe/0002-Fix-gfortran-detection-for-ctng-based-cross-compilers.patch  (That patch was not sent upstream because it breaks other gcc builds. Need to figure out a patch that works for all)

Thanks, I'll have a look.

 

Also, I'm curious to hear more about the illegal instructions. We can fix that if you report them. I'm a conda-forge maintainer.

Here you go, looks like the problem is in Linbox:

https://github.com/conda-forge/sage-feedstock/issues/57

 

We also have Normaliz packaged in conda-forge.

When I try to install "normaliz", mamba wants to role back my version of "sagelib" from 9.1 to 9.0.  I will make a ticket for this too.

Thanks,

Nathan