Re: Presentation and some questions/remarks about numpy/scipy packages
Le dimanche 08 février 2009 à 21:36 -0800, Ondrej Certik a écrit : should build numpy (and scipy) without ATLAS, even if atlas is detected. Thanks for the tip --- I think there was some reason for this conflict, but I can't remember now. Let's wait a bit what other think here. IIRC this is because ATLAS can be installed at runtime if you link against libblas, and the reverse is not true. -- .''`. : :' : We are debian.org. Lower your prices, surrender your code. `. `' We will add your hardware and software distinctiveness to `-our own. Resistance is futile. signature.asc Description: Ceci est une partie de message numériquement signée
Re: Presentation and some questions/remarks about numpy/scipy packages
David Cournapeau schrieb: Hi, I am a developer of numpy and scipy, and I would be interested in helping numpy/scipy debian packages to be better on both Debian and Ubuntu. As a user of both Debian and Ubuntu, and one of the main developer working on build issues for numpy/scipy while not a build issue, the modularity for an install is painful. numpy is too big (including it's dependencies) to be installed on a CD (Ubuntu live CD/alternate CD), however it is required by some other packages, which are itself not packaged in a modular way: pygtk requires for some obscure (couldn't find code using this) the multiarray extension and depends on it (although an exception is thrown at runtime if the extension is not found). while pygtk could be better packaged, you could do the same with numpy by splitting out some sub-packages. Unfortunately splitting out python subpackages is not that easy: Having numpy/core/multiarray.so in it's own binary package would require packaging python-multiarray with the files: numpy/__init__.py numpy/core/__init__.py numpy/core/multiarray.so with the __init__.py files being empty files. Installing on top of this the python-numpy package would require to replace (divert) the __init__.py files with the __init__.py files as found in numpy (packaging tools like dpkg and rpm don't allow overwriting of files without special care). Consider that I additionally might want to split out stuff only needed at runtime (splitting out distutils, tests and include) and legacy stuff (numarray, numeric), I have to add more variants of __init__.py files and maintainance becomes harder. What I really want is to be able to ship independent subpackages as it's own package. Having e.g. the multiarray extension in a subpackage seems to be worse than having it as a toplevel module or in its own namespace. Background in https://bugs.launchpad.net/bugs/309215 Matthias -- To UNSUBSCRIBE, email to debian-python-requ...@lists.debian.org with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Re: Presentation and some questions/remarks about numpy/scipy packages
Le mercredi 11 février 2009 à 22:48 +0100, Matthias Klose a écrit : pygtk requires for some obscure (couldn't find code using this) the multiarray extension and depends on it (although an exception is thrown at runtime if the extension is not found). while pygtk could be better packaged, Suggestions for improvements are always welcome. However, I must have lost the bug report where you explained this. Could you point me to it? Unfortunately splitting out python subpackages is not that easy: Having numpy/core/multiarray.so in it's own binary package would require packaging python-multiarray with the files: numpy/__init__.py numpy/core/__init__.py numpy/core/multiarray.so with the __init__.py files being empty files. Installing on top of this the python-numpy package would require to replace (divert) the __init__.py files with the __init__.py files as found in numpy (packaging tools like dpkg and rpm don't allow overwriting of files without special care). Or you can use python-support which will generate empty __init__.py files if needed, for precisely this kind of use case. -- .''`. : :' : We are debian.org. Lower your prices, surrender your code. `. `' We will add your hardware and software distinctiveness to `-our own. Resistance is futile. signature.asc Description: Ceci est une partie de message numériquement signée
Re: Presentation and some questions/remarks about numpy/scipy packages
On Thu, Feb 12, 2009 at 6:20 AM, Josselin Mouette j...@debian.org wrote: Le dimanche 08 février 2009 à 21:36 -0800, Ondrej Certik a écrit : should build numpy (and scipy) without ATLAS, even if atlas is detected. Thanks for the tip --- I think there was some reason for this conflict, but I can't remember now. Let's wait a bit what other think here. IIRC this is because ATLAS can be installed at runtime if you link against libblas, and the reverse is not true. Yes, that's what I understood. The recent change still links against the general blas/lapack (so that it can be changed at runtime), without the need for build-conflicts. cheers, David -- To UNSUBSCRIBE, email to debian-python-requ...@lists.debian.org with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Re: Presentation and some questions/remarks about numpy/scipy packages
Hi Piotr, On Mon, Feb 9, 2009 at 5:43 PM, Piotr Ożarowski pi...@debian.org wrote: [David Cournapeau, 2009-02-09] I am a developer of numpy and scipy, and I would be interested in helping numpy/scipy debian packages to be better on both Debian and Ubuntu. great! :-) Please create an account on alioth.debian.org and I'll add you to the Debian Python Modules Team[1] so that you'll have commit rights. Thank you, I created an account, cdavid-guest. cheers, David -- To UNSUBSCRIBE, email to debian-python-requ...@lists.debian.org with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Re: Presentation and some questions/remarks about numpy/scipy packages
[David Cournapeau, 2009-02-09] I am a developer of numpy and scipy, and I would be interested in helping numpy/scipy debian packages to be better on both Debian and Ubuntu. great! :-) Please create an account on alioth.debian.org and I'll add you to the Debian Python Modules Team[1] so that you'll have commit rights. If you IRC, you can find us on #debian-python @ OFTC [1] http://python-modules.alioth.debian.org/python-modules-policy.html -- -=[ Piotr Ożarowski ]=- -=[ http://www.ozarowski.pl ]=- pgpMVtpJS2oE7.pgp Description: PGP signature
Presentation and some questions/remarks about numpy/scipy packages
Hi, I am a developer of numpy and scipy, and I would be interested in helping numpy/scipy debian packages to be better on both Debian and Ubuntu. As a user of both Debian and Ubuntu, and one of the main developer working on build issues for numpy/scipy, I think I can be useful :) I looked at the recent debian packaging (current trunk), and I noticed two things which I think are mistakes, or at least seem strange: - build-depends has a depdency on libfftw3-dev. Numpy does not use FFTW at all, and never had to my knowledge, at least from numpy 0.9.8 (almost 3 years ago). - build-conflicts against atlas. I guess this is to avoid explicitely linking against ATLAS - this can be avoided easily without the build-conflicts which is a bit of a pain: ATLAS=None python setup.py build should build numpy (and scipy) without ATLAS, even if atlas is detected. cheers, David -- To UNSUBSCRIBE, email to debian-python-requ...@lists.debian.org with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Re: Presentation and some questions/remarks about numpy/scipy packages
Hi David! On Sun, Feb 8, 2009 at 9:10 PM, David Cournapeau courn...@gmail.com wrote: Hi, I am a developer of numpy and scipy, and I would be interested in helping numpy/scipy debian packages to be better on both Debian and Ubuntu. As a user of both Debian and Ubuntu, and one of the main developer working on build issues for numpy/scipy, I think I can be useful :) I am really glad you are interested in the packaging. Welcome! :) Do you know how to build it from our svn? If not, feel free to ask. Since python-sphinx 0.5 is in unstable now, do you think we should build with the docs as they are in the debian packaging svn? I simply used numpy svn and copied the doc directory into the tarball, see the debian/rules. I looked at the recent debian packaging (current trunk), and I noticed two things which I think are mistakes, or at least seem strange: - build-depends has a depdency on libfftw3-dev. Numpy does not use FFTW at all, and never had to my knowledge, at least from numpy 0.9.8 (almost 3 years ago). I think it could probably be just removed, unless someone objects. - build-conflicts against atlas. I guess this is to avoid explicitely linking against ATLAS - this can be avoided easily without the build-conflicts which is a bit of a pain: ATLAS=None python setup.py build should build numpy (and scipy) without ATLAS, even if atlas is detected. Thanks for the tip --- I think there was some reason for this conflict, but I can't remember now. Let's wait a bit what other think here. Ondrej -- To UNSUBSCRIBE, email to debian-python-requ...@lists.debian.org with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Re: Presentation and some questions/remarks about numpy/scipy packages
On Sun, Feb 8, 2009 at 10:00 PM, David Cournapeau courn...@gmail.com wrote: On Mon, Feb 9, 2009 at 2:36 PM, Ondrej Certik ond...@certik.cz wrote: Hi David! On Sun, Feb 8, 2009 at 9:10 PM, David Cournapeau courn...@gmail.com wrote: Hi, I am a developer of numpy and scipy, and I would be interested in helping numpy/scipy debian packages to be better on both Debian and Ubuntu. As a user of both Debian and Ubuntu, and one of the main developer working on build issues for numpy/scipy, I think I can be useful :) I am really glad you are interested in the packaging. Welcome! :) Do you know how to build it from our svn? If not, feel free to ask. Well, I know how to build a debian package from source and how to fix some issues, but I am not familiar with a lot of tools around. Is there a simple guideline about this ? Our page is here: http://wiki.debian.org/Teams/PythonModulesTeam and in particular I wrote this howto: http://wiki.debian.org/PAPT_Howto E.g. the section Building the package. Btw, feel free to update the PythonModulesTeam page above with instructions how to build the package, that would be really useful. Since python-sphinx 0.5 is in unstable now, do you think we should build with the docs as they are in the debian packaging svn? I simply used numpy svn and copied the doc directory into the tarball, see the debian/rules. Doc is a pain to deal with with distutils, generally. I don't know what's the doc status in 1.2 (the documentation infrastructure is one of the big change since 1.2 for the upcoming 1.3). I think we can ignore doc changes for 1.2*, and focus on 1.3, since we would need to change anyway for 1.3. Ok, so should I leave the current tarball, as you can see, I am using svn -r6221: get-orig-source: wget http://qa.debian.org/watch/sf.php/numpy/numpy-1.2.1.tar.gz; \ tar xzf numpy-1.2.1.tar.gz; \ cd numpy-1.2.1; \ svn export -r6221 http://svn.scipy.org/svn/numpy/trunk/doc; \ cd ..; \ tar czf python-numpy_1.2.1.ds.orig.tar.gz numpy-1.2.1; \ rm -rf numpy-1.2.1 numpy-1.2.1.tar.gz Thanks for the tip --- I think there was some reason for this conflict, but I can't remember now. Let's wait a bit what other think here. According to the debian/changelog file and the related bug on debian bug tracker, this was to avoid depending on ATLAS at link stage. This can be done by the command above - which is not really documented. Generally, the numpy build system is a bit arcane and 'magic'. In the short term, I hope to clarify most of oddities for easier packaging (not just Debian, BTW). Thanks, that's excellent. Ondrej -- To UNSUBSCRIBE, email to debian-python-requ...@lists.debian.org with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
