tree 754340ca3aa6a589f01561460ee219397ebb1c0b
parent 72c8f564e1fef0b93076d152869fd860b00a7399
author bjacob <benoitjacob@google.com> 1684342350 -0400
committer GitHub <noreply@github.com> 1684342350 +0000
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJkZQZOCRBK7hj4Ov3rIwAAG0cIAAou4HnGkWHZHKeB8wWwG/6r
 MvSZflGpYCTfwnxWdTRP8DUXUIOuYOLF6TR4Mf5nOt7nhd2WQZZWNxWzc7mLiMf5
 +PTV/Dz5Ot9xEmhU9Pxad0a9aCCk9RugZdN7KzlWXIX/sS7SvRQQ1Pnbqf2h6xHc
 umghft4wgZwKSwM81t/nf6uREY2kulpvP3O9CP/3SChzd15FNYB551jcgfkNPSCQ
 zP8Zt1SJZCWh1F+f5UlatB4x2ttS8tcgXGP7RCvq2I9EcrMDgqumeSflP/LNntAE
 axUduqVn9U8dS2xW2mcx0qWE5OlaLZwZSjlRRxUDSiIDwG0IpwapPI6JBe4LZ0M=
 =loUi
 -----END PGP SIGNATURE-----
 

e2e matmul test improvements (#13657)

* Test the non-accumulating matmul case (fill-zero accumulator). See
discussion in #13582 with @ThomasRaoux .
* Drop the `MIXED` dynamicity: it was already unused, having been
dropped out of actual test variants iterated over. Drop helpers that
were only used for it, such as `pseudorandom_bool`.
* In case of numerical error, drop the logic re-running matmul tests on
simpler matrices. Although potentially useful to help pinpoint the
numerical issue, it was confusing when debugging (as discussed once with
@manishucsd ), and it made test runner code substantially more complex.