[petsc-users] Libraries destination directory is non-standard

[Juan Luis Cano]

On 2015-01-11 18:41, Lisandro Dalcin wrote:
> On 8 January 2015 at 23:32, Juan Luis Cano <juanlu001 at gmail.com> wrote:
>> Hello all,
>>
>> I am trying to build a conda package of PETSc 3.4 using its Python build
>> system so anybody can install it in the Anaconda Python distribution. I need
>> PETSc as a dependency for FEniCS.
>>
> Nice to know! I was thinking about doing it myself. You may find
> interesting my mpich and openmpi packages (newer versions than the
> available in Anaconda) to depend on https://binstar.org/mpi4py, I
> still have to put the recipes in github repo.
>
>> The problem is that the required shared libraries (also those corresponding
>> to external packages such as UMFPACK) are installed in
>> $PREFIX/lib/python2.7/site-packages/petsc/lib. The build process goes
>> perfectly fine but when packaging this into a tarball these libs won't be
>> found by any package as they are not in $PREFIX/lib. Here is the build
>> script I am using:
>>
> Are you using pip to create/build the conda recipes? I really advice
> against that. You should create a petsc package that is independent of
> a Python runtime (of course, you still need Python to run configure),
> this way you can install in the standard $PREFIX location. Why do you
> say things are installed in site-packages? Who is installing there
> FEniCS ?
¡Hola Lisandro! Thanks for your answer. At first I used 'conda skeleton 
pypi' to create the packages, but the setup.py procedure installs in 
site-packages and it didn't seem appropriate, so I finally arranged the 
build script as you say.

> Please note/remember that conda is not only for Python stuff, but for
> any other Python-independent package that install under a $PREFIX
> tree.

I am aware! Actually if you see my binstar you will see boost and eigen3 
there.

>> https://github.com/Juanlu001/conda-recipes/blob/juanlu001/fenics/petsc/build.sh
>>
> * Please remember to add the md5 hash under the source section.
> * python should not be a runtime dependency of the petsc package

You are right, I wanted to do it the quick 'n' dirty way so I didn't 
write the hashes. This evening I achieved the result I wanted so I will 
polish the recipes.

>> I tried to change the destination using `--prefix` in
>> PETSC_CONFIGURE_OPTIONS and after `setup.py install`, but neither worked.
>> Changing LD_LIBRARY_PATH seems dangerous to me and creating symbolic links
>> three levels higher seems dirty. If the python version changes, or "lib64"
>> is added somewhere, this will surely break... Does anybody have a suggestion
>> on how to do this?
>>
> Knowing the full stack of conda packages you depend on would help to
> figure out your issue.

As I said, fortunately this is not an issue anymore since I am not using 
'setup.py install'. The configure && make procedure is much more suitable.

Regards,

Juan Luis