[Numpy-discussion] Switch to using ATLAS for OSX binary wheels
Matthew Brett
matthew.brett at gmail.com
Mon Jun 16 08:00:53 EDT 2014
Hi,
On Mon, Jun 16, 2014 at 12:51 PM, <josef.pktd at gmail.com> wrote:
> On Mon, Jun 16, 2014 at 7:10 AM, Matthew Brett <matthew.brett at gmail.com> wrote:
>> Hi,
>>
>> On Sun, Jun 15, 2014 at 10:51 PM, Ralf Gommers <ralf.gommers at gmail.com> wrote:
>>>
>>>
>>>
>>> On Sat, Jun 14, 2014 at 11:56 AM, Matthew Brett <matthew.brett at gmail.com>
>>> wrote:
>>>>
>>>>
>>>>
>>>> On Friday, June 13, 2014, Ralf Gommers <ralf.gommers at gmail.com> wrote:
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On Fri, Jun 13, 2014 at 2:07 PM, Matthew Brett <matthew.brett at gmail.com>
>>>>> wrote:
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> Summary : I'm planning to upload OSX wheels for numpy and scipy using
>>>>>> the ATLAS blas / lapack library instead of the default OSX Accelerate
>>>>>> framework.
>>>>>>
>>>>>> We've run into some trouble with a segfault in recent OSX Accelerate:
>>>>>>
>>>>>> https://github.com/numpy/numpy/issues/4007
>>>>>>
>>>>>> and Accelerate also doesn't play well with multiprocessing.
>>>>>>
>>>>>> https://github.com/numpy/numpy/issues/4776
>>>>>>
>>>>>> Because there's nothing I love more than half-day compilation runs on
>>>>>> my laptop, I've built ATLAS binaries with gcc 4.8, and linked numpy
>>>>>> and scipy to them to make OSX wheels. These pass all tests in i386
>>>>>> and x86_64 mode, including numpy, scipy, matplotlib, pandas:
>>>>>>
>>>>>>
>>>>>> https://travis-ci.org/matthew-brett/scipy-stack-osx-testing/builds/27442987
>>>>>>
>>>>>> The build process needs some automating, but it's recorded here:
>>>>>>
>>>>>> https://github.com/matthew-brett/numpy-atlas-binaries
>>>>>>
>>>>>> It's possible to get travis-ci to build these guys from a bare machine
>>>>>> and then upload them somewhere, but I haven't tried to do that.
>>>>>>
>>>>>> Meanwhile Sturla kindly worked up a patch to numpy to work round the
>>>>>> Accelerate segfault [1]. I haven't tested that, but given I'd already
>>>>>> built the wheels, I prefer the ATLAS builds because they work with
>>>>>> multiprocessing.
>>>>>>
>>>>>> I propose uploading these wheels as the default for numpy and scipy.
>>>>>> Does anyone have any objection or comments before I go ahead and do
>>>>>> that?
>>>>>
>>>>>
>>>>> From your README and wscript I don't see what numpy version you're using
>>>>> to compile scipy against. I got the impression that you used 1.8.1, but it
>>>>> should be numpy 1.5.1 for the 2.7 build, and 1.7.1 for 3.x.
>>>>>
>>>>> I've tried the scipy 0.14.0 python2.7 wheel, but I get import errors (see
>>>>> below). Your wheels should work with all common Python installs (mine is
>>>>> homebrew) right?
>>>>>
>>>>> Ralf
>>>>>
>>>>>
>>>>> $ python2.7 -c "import scipy; scipy.test()"
>>>>> Running unit tests for scipy
>>>>> NumPy version 1.9.0.dev-056ab73
>>>>> NumPy is installed in
>>>>> /usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/numpy
>>>>> SciPy version 0.14.0
>>>>> SciPy is installed in
>>>>> /usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/scipy
>>>>> Python version 2.7.5 (default, Jun 18 2013, 21:21:44) [GCC 4.2.1
>>>>> Compatible Apple LLVM 4.2 (clang-425.0.28)]
>>>>> nose version 1.3.0
>>>>> E...............EEEEEE............EEEEEEEEEE
>>>>>
>>>>> ======================================================================
>>>>> ERROR: Failure: ImportError
>>>>> (dlopen(/usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/scipy/cluster/_hierarchy_wrap.so,
>>>>> 2): Symbol not found: _PyModule_Create2
>>>>> Referenced from:
>>>>> /usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/scipy/cluster/_hierarchy_wrap.so
>>>>> Expected in: flat namespace
>>>>> in
>>>>> /usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/scipy/cluster/_hierarchy_wrap.so)
>>>>>
>>>>> ----------------------------------------------------------------------
>>>>> Traceback (most recent call last):
>>>>> File
>>>>> "/usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/nose/loader.py",
>>>>> line 413, in loadTestsFromName
>>>>> addr.filename, addr.module)
>>>>> File
>>>>> "/usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/nose/importer.py",
>>>>> line 47, in importFromPath
>>>>> return self.importFromDir(dir_path, fqname)
>>>>> File
>>>>> "/usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/nose/importer.py",
>>>>> line 94, in importFromDir
>>>>> mod = load_module(part_fqname, fh, filename, desc)
>>>>> File
>>>>> "/usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/scipy/cluster/__init__.py",
>>>>> line 27, in <module>
>>>>> from . import vq, hierarchy
>>>>> File
>>>>> "/usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/scipy/cluster/hierarchy.py",
>>>>> line 175, in <module>
>>>>> from . import _hierarchy_wrap
>>>>> ImportError:
>>>>> dlopen(/usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/scipy/cluster/_hierarchy_wrap.so,
>>>>> 2): Symbol not found: _PyModule_Create2
>>>>> Referenced from:
>>>>> /usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/scipy/cluster/_hierarchy_wrap.so
>>>>> Expected in: flat namespace
>>>>> in
>>>>> /usr/local/Cellar/python/2.7.5/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/scipy/cluster/_hierarchy_wrap.so
>>>>>
>>>>> ...
>>>>> <42 more errors>
>>>>
>>>>
>>>> That is strange homebrew is one of tests in the grid and the installation
>>>> path looks strange.
>>>>
>>>> Can you try downloading the wheel from the url and installing from the
>>>> local file?
>>>
>>>
>>> That's what I did (easier than remembering the magic pip incantation). The
>>> install path looks fine to me, maybe homebrew changed it recently? I can try
>>> to update my install, will take a few days though.
>>
>> Yes, sorry - that does look like the normal homebrew install path, I
>> didn't realize it had the exotic framework parts to it.
>>
>> I just ran these commands on my machine:
>>
>> SPI=https://nipy.bic.berkeley.edu/scipy_installers
>> BREW_BIN=/usr/local/Cellar/python/2.7.6/bin
>> curl -O $SPI/numpy-1.8.1-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.whl
>> curl -O $SPI/scipy-0.14.0-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.whl
>> $BREW_BIN/pip install
>> numpy-1.8.1-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.whl
>> $BREW_BIN/pip install
>> scipy-0.14.0-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.whl
>> $BREW_BIN/python -c "import scipy; scipy.test()"
>>
>> and the scipy tests passed.
>>
>> I built the scipy wheel against numpy 1.8.1 - but - aren't the numpies
>> binary compatible? What difference would I expect to see if I'd built
>> scipy against numpy 1.5.1 or 1.7 ?
>
> I summarized here what David explained a while ago about the
> difference between forward and backwards binary compatibility
>
> http://stackoverflow.com/questions/17709641/valueerror-numpy-dtype-has-the-wrong-size-try-recompiling/18369312#18369312
>
I see - thanks for the summary. I will recompile.
Cheers,
Matthew
More information about the NumPy-Discussion
mailing list