tree 3afa9beb3242d6696a77fc45f7f6801ecc311bbf
parent 8c342c75a8f6aa1128a5e34d34c96b2ea68b8f1b
author bjacob <benoitjacob@google.com> 1680317130 -0400
committer GitHub <noreply@github.com> 1680317130 +0000
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJkJ5rKCRBK7hj4Ov3rIwAA8PgIAEv68oAQbTJ5YIP6FQECsoqM
 emGfI6IcnDpDwbd+e/8wljcguqX72MvDJH9aYD1DdEvUpgEUcAD8ZT0oGGRivzeT
 am47qNtCNzoIyy74MEG13V7rGKJ2ZutEaWxZn1VVp35TddLk+2NZqS8MM+pkmOBY
 9uSemR6lsGhCVmv9pB1OgR0FVNtoUWr8QumexNTwW6Ei5Wf6UhfViX4rwNpFoI6R
 uwKYIEhYs+avln0yoii53gzaVBuVJkirnKYsyVv5yzLi9A0KSySzONDBEJRUOR/D
 T66p4LkeSGCDKOaap0i+PtLTWsc9HrvnKTYIHJLZ5U6po9JKeGiJ/WBXu63NOTU=
 =Apla
 -----END PGP SIGNATURE-----
 

Avoid more sanitizer test timeouts (#12887)

More CI timeouts with TSAN in `pack_test` reported by @stellaraccident :

https://github.com/openxla/iree/actions/runs/4580136815/jobs/8088573298?pr=12884

This PR shrinks TSAN test latency on my desktop from 12 seconds to 0.85
second. I think it preserves the really important test coverage still.
The sole substantially large remaining shape `{9, 33}` is larger in
dimension 1 because that is the inner loop dim and that's where we want
to cast the most bugs. Values are just above a power of two. And the
64->16 change is because in practice when we do "pad a lot" that is
currently incidentally due to our pad-everything-by-16x16 approach to
make buffer allocations independent of materialize-encoding, so the 16
value is what we care about in practice.