Re: [SciPy-Dev] [Numpy-discussion] ANN: Scipy 0.13.0 beta 1 release
On Mon, Sep 2, 2013 at 5:46 PM, Pauli Virtanen <pav@iki.fi> wrote:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Hi,
02.09.2013 18:30, David Cournapeau kirjoitti: [clip]
Ok, I think the commit 5935030f8cced33e433804a21bdb15572d1d38e8 is quite wrong.
It conflates the issue of dealing with Accelerate brokenness and using g77 ABI. I would suggest reverting it for 0.13.0 (and re-disable single precision), as fixing this correctly may require quite some time/testing.
I'm -1 on returning to the previous situation where many routines on OSX are simply broken (which was the situation previously --- "disabling single precision" left several things still broken). I'd rather just postpone the 0.13.0 release until this issue is solved properly.
I see, I missed that it was more than just reverting to slower versions.
Can you say what exactly is wrong -- as far as I know, Accelerate on OSX simply uses g77 ABI. There were some bugs previously where it in places did things differently, but on recent OSX releases this is no longer the case?
I guess you are trying to link with MKL? In that case, I would rather propose restoring the previous MKL wrappers (with trivial extensions for the missing w* symbols).
The pb is specific to MKL, yes, but I've found a workaround: in the case of MKL, you just need to workaround the g77 ABI of the MKL vs gfortran ABI, but the LAPACK interface is the actual LAPACK, not CLAPACK. So for MKL, you need non dummy wrappers iff the function returns a complex. A dirty patch seems to confirm that this fixes the issue, I will prepare an actual patch tomorrow, David
- -- Pauli Virtanen -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.12 (GNU/Linux)
iEYEARECAAYFAlIkwPQACgkQ6BQxb7O0pWBnkwCfeKj32CYYCPdEWVcYYMtq/OZM 32wAoIn9yG/HNtUeh+XwqAm2uAS9sVQ5 =CKBN -----END PGP SIGNATURE-----
_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Hi, 02.09.2013 21:12, David Cournapeau kirjoitti: [clip]
The pb is specific to MKL, yes, but I've found a workaround: in the case of MKL, you just need to workaround the g77 ABI of the MKL vs gfortran ABI, but the LAPACK interface is the actual LAPACK, not CLAPACK.
So for MKL, you need non dummy wrappers iff the function returns a complex. A dirty patch seems to confirm that this fixes the issue, I will prepare an actual patch tomorrow,
Thanks, that sounds like an OK solution. I don't see a problem having separate wrappers for MKL; it's sort of a maintenance burden, but this probably can't be helped as the ABIs don't sound like they are compatible. The new ABI wrappers were obviously not tested against MKL, so we have to do this now... Best, Pauli -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.12 (GNU/Linux) iEYEARECAAYFAlIk2ZoACgkQ6BQxb7O0pWDUKwCfR35y4AS2zECzr2y7enwaPEbH JUoAoKEG1ptACEiuShbhr3Tg/QtrpFSz =irId -----END PGP SIGNATURE-----
On Mon, Sep 2, 2013 at 7:31 PM, Pauli Virtanen <pav@iki.fi> wrote:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Hi,
02.09.2013 21:12, David Cournapeau kirjoitti: [clip]
The pb is specific to MKL, yes, but I've found a workaround: in the case of MKL, you just need to workaround the g77 ABI of the MKL vs gfortran ABI, but the LAPACK interface is the actual LAPACK, not CLAPACK.
So for MKL, you need non dummy wrappers iff the function returns a complex. A dirty patch seems to confirm that this fixes the issue, I will prepare an actual patch tomorrow,
Thanks, that sounds like an OK solution. I don't see a problem having separate wrappers for MKL; it's sort of a maintenance burden, but this probably can't be helped as the ABIs don't sound like they are compatible.
The new ABI wrappers were obviously not tested against MKL, so we have to do this now...
See https://github.com/scipy/scipy/pull/2824 I decided to split our wrappers into the G77 ABI-specific part (returning complex) and the CBLAS/CLAPACK hacks specific to Accelerate. Not sure the complexity is worth it compared to simply have three cases (MKL on OS X, Accelerate, rest), but it may help integration with other libraries. David
Best, Pauli -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.12 (GNU/Linux)
iEYEARECAAYFAlIk2ZoACgkQ6BQxb7O0pWDUKwCfR35y4AS2zECzr2y7enwaPEbH JUoAoKEG1ptACEiuShbhr3Tg/QtrpFSz =irId -----END PGP SIGNATURE-----
_______________________________________________ SciPy-Dev mailing list SciPy-Dev@scipy.org http://mail.scipy.org/mailman/listinfo/scipy-dev
participants (2)
-
David Cournapeau -
Pauli Virtanen