[Bug 1177260] Switching python-numpy to plain BLAS causes significant performance regression
https://bugzilla.suse.com/show_bug.cgi?id=1177260 https://bugzilla.suse.com/show_bug.cgi?id=1177260#c38 --- Comment #38 from Stefan Br�ns <stefan.bruens@rwth-aachen.de> --- (In reply to Egbert Eich from comment #37)
Stefan Br�ns from comment #36)
Unfortunately for Leap, the current python3-numpy package is crippled again, as the cblas dependency has been removed.
So while we are likely good on TW, the funny smell of SLE/Leap 15 becomes hard to ignore.
This is pretty much a catch-22 situation. We need to have cblas available in SLE before we can build numpy with it. It used to be built against openblas, but this made openblas a hard runtime requirement for it. If it is built with a version of netlib-lapack, that provides cblas, you can optionally use OpenBLAS as well. SO we will first have to provide CBLAS and then we can fix up numpy.
It is still available in the backports, e.g. https://build.opensuse.org/package/show/openSUSE:Backports:SLE-15-SP4/cblas On the other hand, doing anything conditionally for Leap/SLE 15 in the current python-numpy package is quite pointless, as it requires Python 3.8.
(And no, missing cblas is not "unfixable", Leap 15.2 still had it.)
Are you sure? What change made it go away? Wasn't it the switch of numpy to netlib-lapack that caused this?
CBLAS was imported into LAPACK in August 2014 [1], and became part of LAPACK 3.6.0 (released November 2015 [2]). Leap 15.2 had the standalone CBLAS (which is just a fairly thin wrapper around BLAS), which is still available here: https://netlib.org/blas/#_cblas [1] https://github.com/Reference-LAPACK/lapack-release/blob/8d160e5f960b389ba6a4... [2] https://github.com/Reference-LAPACK/lapack/blob/7866626840f5d5e7e27f027a5518... -- You are receiving this mail because: You are on the CC list for the bug.
participants (1)
-
bugzilla_noreply@suse.com