[Numpy-discussion] Scipy build can't find BLAS when using numpy master? (Was: Should arr.diagonal() return a copy or a view? (1.7 compatibility issue))

Nathaniel Smith njs at pobox.com
Wed May 16 13:52:33 EDT 2012


On Wed, May 16, 2012 at 4:50 PM, Robert Kern <robert.kern at gmail.com> wrote:
> On Wed, May 16, 2012 at 4:35 PM, Nathaniel Smith <njs at pobox.com> wrote:
>> On Wed, May 16, 2012 at 4:24 PM, Robert Kern <robert.kern at gmail.com> wrote:
>>> On Wed, May 16, 2012 at 4:21 PM, Nathaniel Smith <njs at pobox.com> wrote:
>>>
>>>> I built some pristine python 2.7 installs from scratch (no virtualenv,
>>>> no distro tweaks, etc.). Then I installed some version of numpy in
>>>> each, then tried building scipy. With numpy 1.6.1 (built from git),
>>>> everything seems fine - it finds atlas in /usr/lib64/atlas.
>>>>
>>>> With current numpy master (3bbbbd416a0a), numpy itself builds fine,
>>>> but scipy's setup.py can't find atlas (even though it's looking in the
>>>> right places). Log attached.
>>>
>>> The log says it's looking in /usr/lib64/atlas-base/, not /usr/lib64/atlas/.
>>
>> Sorry, that was an error in my email. The libraries are actually in
>> /usr/lib64/atlas-base, and when I build against numpy 1.6.1, the build
>> log says:
>>
>> Setting PTATLAS=ATLAS
>>  FOUND:
>>    libraries = ['ptf77blas', 'ptcblas', 'atlas']
>>    library_dirs = ['/usr/lib64/atlas-base']
>>    language = c
>>    define_macros = [('ATLAS_INFO', '"\\"3.8.3\\""')]
>>    include_dirs = ['/usr/include/atlas']
>>
>>  FOUND:
>>    libraries = ['ptf77blas', 'ptcblas', 'atlas']
>>    library_dirs = ['/usr/lib64/atlas-base']
>>    language = c
>>    define_macros = [('ATLAS_INFO', '"\\"3.8.3\\""')]
>>    include_dirs = ['/usr/include/atlas']
>
> Hmm. I would throw in some print statements into the "_check_libs()"
> and "_lib_list()" methods in numpy/distutils/system_info.py and see
> what it's checking for.

It turns out that len(found) == len(expected) is not a good test in
situations where you might find *more* than you expected...

https://github.com/numpy/numpy/pull/281

- N



More information about the NumPy-Discussion mailing list