On Mon Mar 25, 2024 at 4:47 PM CET, Ralf Gommers wrote: > > Is there anything more to do than just switch of particular features in > > `--cpu-dispatch`? > > I suspect not, assuming that that works. If Z13 support is an important use > case where the performance difference matters a lot, maybe you can bring up > getting partial or even full support back in NumPy. Most NumPy maintainers > won't know much/anything about s390x though, so that may require > contributions and be nontrivial. > > Another, more feasible option would be to auto-disable the SIMD features, > emit a warning, and then continue rather than fail the NumPy build on Z13 > by default.
My colleague Fabian Vogt made a deep debug and found this (see ): # I found that the test failure here was also # excluded on 32bit ARM and I dug a bit further: # https://github.com/numpy/numpy/issues/24548 shows the same reason # as here. That's when I also noticed the additional info at the # bottom: # # feats = 'VXE, None' # # At first I wasn't able to reproduce it on armv7 because cpu # feature detection was broken, which nobody noticed... I debugged # that and fixed it with https://github.com/numpy/meson/pull/12. # # After fixing that, I was able to reproduce the test # failure locally by using NPY_ENABLE_CPU_FEATURES="ASIMDHP, # None". This wasn't meant to succeed but it did. Fixed with # https://github.com/numpy/numpy/pull/26151. # # Both submitted as # https://build.opensuse.org/request/show/1163337 Best, Matěj -- http://matej.ceplovi.cz/blog/, @mcepl@floss.social GPG Finger: 3C76 A027 CA45 AD70 98B5 BC1D 7920 5802 880B C9D8 When God put a calling on your life, He already factored in your stupidity. -- Unknown, mentioned in a sermon
E09FEF25D96484AC.asc
Description: application/pgp-keys
signature.asc
Description: PGP signature
_______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com