[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#c43 --- Comment #43 from Egbert Eich <eich@suse.com> --- (In reply to Stefan Br�ns from comment #41)
1. Build numpy as part of Backports, which already *has* cblas. 2. Bring an updated lapack (>= 3.6.0) to SLE. But as far as I understand, this *also* goes into backports only, so we also need to build numpy in backports.
We're actually pursuing option 2. lapack has been fighting us all the way, on SLE/Leap we suddenly saw errors in the tests due to precision issues. I'm in the process of fixing this by setting -mfpmath=sse there. On Factory, we cannot do this so we leave -ffloat-store for the tests. With a future compiler this may well break as precision using the 80bit long-double extended precision in the fpu may always cause this. -- You are receiving this mail because: You are on the CC list for the bug.
participants (1)
-
bugzilla_noreply@suse.com