<div dir="ltr">On Thu, Apr 17, 2014 at 8:29 AM, Grant Goodyear <span dir="ltr">&lt;<a href="mailto:g2boojum@gmail.com" target="_blank">g2boojum@gmail.com</a>&gt;</span> wrote:<br><div class="gmail_extra"><div class="gmail_quote">
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div><div>For reproducibility, I like to have a python virtualenv that has all of the python libraries that I use, relying on portage just for python, virtualenv, and virtualenvwrapper. Building numpy using &quot;pip install numpy&quot; fails if I use the science-overlay versions of the reference lapack and blas libraries, because pip can&#39;t find liblapack to link to it.  <br>

<br></div>Right now I&#39;m cheating and using the non-overlay versions, but I&#39;m sure that somebody has a better solution of how to help pip find the right libraries.  Help?<br></div></div></blockquote><div><br></div>
<div style>The problem lies in the broken design of numpy.distutils trying to do way too much autodetection, and our ways to manage around it while keeping our linear algebra modularity.</div><div style>A possible workaround if you really want to use pip and the gentoo-science overlay framework is to simply link your selected blas and lapack libraries to /usr/lib/lib{blas,lapack}.so. I have not tested it, but I suspect this should work for openblas and  {blas,lapack}-reference since they only have one library to link.</div>
<div style><br></div><div style>Are you trying to do cross-distro reproducibility? </div><div style><br></div><div style>Sebastien</div><div> </div></div><br></div></div>