Hi, On 07/04/21 at 10:56 +0200, Andreas Tille wrote: > Hi Lucas, > > On Wed, Apr 07, 2021 at 08:49:27AM +0200, Lucas Nussbaum wrote: > > I confirm that I can still reproduce this failure in a testing chroot. > > Thank you for your information. I have tried to create a diff between > the successful build log of the autobuilders[1] where all architectures > have passed and your build log snippet. Theses asserts in your build > log seem to cause the build failure: > > ASSERT:expected:<3088458213> but was:<3088444823> > ASSERT:expected:<3088496565> but was:<3088483525> > ASSERT:expected:<338154001.76> but was:<338144655.90> > ASSERT:expected:<15465346.22> but was:<15464740.32> > > This looks pretty much like floating point rounding issues. The > interesting question for me is: If we do not see these assertions in a > variety of architectures with different floating point implementations > what might make your test machine so special to fail in contrast to > the Debian autobuilders. > > If upstream (in CC) has no better idea the only way I personally see is > to comment out those tests that are featuring the expected numbers > above. But it would be better to investigate this issue with your help > in more detail. So if your hardware is in some severe way different > than Debian autobuilders (either in floating point implementation or > possibly in parallelisation - sometimes we see weird things due to > massive parallelisation) it would be great if you could share this > information in this bug report.
I don't think there's anything very specific about the hardware I'm using (Amazon EC2 VMs). I can provide access to those VMs if that's useful (but I need to start them specifically for this test). I've just done a full rebuild of bullseye and looked at all remaining failures, and this is the only issue of that kind. Lucas
