Atlas 3.10 and Numpy

General support questions
Post Reply
jeffng
Posts: 2
Joined: 2014/09/29 18:39:08

Atlas 3.10 and Numpy

Post by jeffng » 2014/09/29 19:00:16

Hello everybody,

First time poster here and I'd like to say great job on CentOS 7.

I am trying to compile Numpy 1.9.0 (via pip) with Atlas support (Python 3.4 installed with conda) but unfortunately, I can't find the usual suspects which were present in CentOS 6.5.

When I try "pip install numpy", I get:

Code: Select all

atlas_blas_threads_info:
    Setting PTATLAS=ATLAS
      libraries ptf77blas,ptcblas,atlas not found in ['/opt/auxiliary-software/lib', '/usr/local/lib64', '/usr/local/lib', '/usr/lib64/atlas', '/usr/lib64/sse2', '/usr/lib64', '/usr/lib/sse2', '/usr/lib', '/usr/lib/sse2', '/usr/lib/']
      NOT AVAILABLE

    atlas_blas_info:
      libraries f77blas,cblas,atlas not found in ['/opt/auxiliary-software/lib', '/usr/local/lib64', '/usr/local/lib', '/usr/lib64/atlas', '/usr/lib64/sse2', '/usr/lib64', '/usr/lib/sse2', '/usr/lib', '/usr/lib/sse2', '/usr/lib/']
      NOT AVAILABLE
On the CentOS 7 atlas and atlas-devel rpms, there are only libsatlas.so and libtatlas.so, whereas on CentOS 6.5, there were libatlas.so, libclbas.so and libf77blas.so, which Numpy seems to be looking for.

Can anybody help, please?

Thanks,
Jeff

chemal
Posts: 776
Joined: 2013/12/08 19:44:49

Re: Atlas 3.10 and Numpy

Post by chemal » 2014/09/30 01:20:31

Using the latest and greatest from other people doesn't always mean that they are also using the latest and greatest themselves. In other words: C7 is too recent. The numpy hackers haven't yet adopted atlas-3.10. But it will compile with reference blas/lapack. You just have to install that.

jeffng
Posts: 2
Joined: 2014/09/29 18:39:08

Re: Atlas 3.10 and Numpy

Post by jeffng » 2014/09/30 20:22:11

In case it is useful for anybody attempting this, ATLAS 3.10 has started to compile all three separate libraries (libatlas.so, libcblas.so and libf77blas.so) into a "fat" lib, i.e. libsatlas.so.

I got round the problem by creating symlinks for those three libraries and I also added some for libptatlas.so, libptcblas.so and libptf77blas.so to libtatlas.so.

Now numpy is happy and dot product is sped up by 300% (although using 4 threads).

Post Reply