[Numpy-discussion] problems building numpy with ACML blas/lapack

[Thomas Unterthiner]

Hi!

(I was contacted about this off-list, but thought it might be worth it 
to document this on-list)

I don't remember ever running into problems while using ACML in my numpy 
installs. However I never ran the whole test-suite, and I never used 
np.einsum (or complex numbers) in my own code. All I can tell you is 
that while using numpy with ACML, I never ran into any weird segfaults.

For what it's worth: OpenBLAS has added a lot of AMD-specific 
(Bulldozer- and Piledriver-specific, to be exact) code after its 0.2.8 
version, which made it slightly faster than ACML on my own machines (at 
least for dgemm, I remember some LAPACK calls were a tiny bit slower, 
but not so much that it bothered me). Since numpy-integration of 
OpenBLAS is significantly better than ACML, I don't bother with the 
latter anymore.

You can download the OpenBLAS 0.2.9rc1 -- which includes the 
Bulldozer/Piledriver code -- from 
https://github.com/xianyi/OpenBLAS/releases/tag/v0.2.9.rc1


Cheers


Thomas





On 2014-02-24 20:58, Michael Hughes wrote:
> Hello,
>
> I'm trying to build numpy from source to use AMD's ACML for matrix 
> multiplication (specifically the multi-threaded versions 
> gfortran64_mp). I'm able to successfully compile and use a working 
> version of np.dot, but my resulting installation doesn't pass numpy's 
> test suite, instead, I get a segfault. I'm hoping for some advice on 
> what might be wrong...
>
>  I'm on Debian, with a fresh install of Python-2.7.6. To install 
> numpy, I've followed exactly the instructions previously posted to 
> this list by Thomas Unterthiner. See 
> http://numpy-discussion.10968.n7.nabble.com/numpy-ACML-support-is-kind-of-broken-td35454.html. 
> The only thing I've adjusted is to try to use the gfortran64_mp 
> version of ACML instead of just gfortran64.
>
> Using those instructions, I can compile numpy-1.8.0 so that it 
> successfully uses the desired ACML libraries.  I can confirm this by 
> `ldd site-packages/numpy/core/_dotblas.so`, which shows that I'm 
> linked to libacml_mp.so as desired. Furthermore, some quick timing 
> tests indicate that for a 1000x1000 matrix X, calls to np.dot(X,X) 
> have similar speeds as using custom C code that directly calls the 
> ACML libraries. So, dot seems to work as desired.
>
> However, when I run numpy.test(verbose=4), I find that I get a seg fault
> ```
> test_einsum_sums_cfloat128 (test_einsum.TestEinSum) ... Segmentation fault
> ```
>
> Any ideas what might be wrong? From my benchmark tests, ACML is way 
> faster than MKL or other options on my system, so I'd really like to 
> use it, but I don't trust this current install.
>
> Thanks!
> - Mike
>
>
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20140227/1855447a/attachment.html>