search

ucberkeley / bce

Berkeley Common Environment provides a common Linux computational environment for classwork and research.
Apache License 2.0
13 stars 6 forks source link

make sure Python links against openBLAS #67

Open paciorek opened 8 years ago

paciorek commented 8 years ago

Not sure how to do this without a bit of research, but we should try to have Miniconda numpy/scipy use openBLAS. The Anaconda page suggests it is using ATLAS.

Perhaps not a must-have, but if we're serious about having BCE set up for scientific computing, we should do this soon.

ryanlovett commented 8 years ago

We'll probably have to conda build numpy/scipy via the -openblas recipes at https://github.com/conda/conda-recipes. If this takes too long we might need to do this outside of bootstrap and host the builds at binstar.

Ryan

On Wed, Sep 23, 2015 at 5:37 PM, Christopher Paciorek < notifications@github.com> wrote:

Not sure how to do this without a bit of research, but we should try to have Miniconda numpy/scipy use openBLAS. The Anaconda page suggests it is using ATLAS.

Perhaps not a must-have, but if we're serious about having BCE set up for scientific computing, we should do this soon.

— Reply to this email directly or view it on GitHub https://github.com/ucberkeley/bce/issues/67.

paciorek commented 8 years ago

I did a brief amount of testing and the current dot (matrix multiply) in Python in BCE-summer-2015 is actually quite fast (surprisingly so considering other comparisons of different BLAS's that I've done in the past). Not threaded though, so still worth linking against openBLAS.

I'll note that dot in Python in BCE-spring-2015 is very slow.

paciorek commented 8 years ago

@ryanlovett: looks like numpy's _dotblas.so (which does dot product/matrix multiply - not sure what else) is now linking against openblas in BCE-2015-fall but not a bunch of numpy's linear algebra. We should chat further about this.