On 12/20/23 20:30, juzhe.zh...@rivai.ai wrote:
OK. Sounds reasonable.
But from my side, I used to commit patches after full coverage testing.
Understood. And it's appreciated -- the code you're doing hits a wide
variety of configurations, so the wider testing is probably applicable.
Ideally the thead vector bits need reasonable testing to make sure they
don't totally break the standard RVV support. So for something like the
final scheme to add the "th." prefix I'd expect they can get away with
just rv64gcv. WHile there is a chance that'll miss something, the odds
are pretty low that a bug will be uncovered for each additional
configuration tested beyond the first.
In contrast if they needed to make a structural changes that are more
than adding a path for thead's vector unit, then we might reasonably ask
for a deeper test of that specific patch (perhaps even suggesting the
configurations most likely affected and thus which need to be tested).
The key being we want to use time wisely and testing dozens of multilibs
for each change isn't really reasonable.
It's always a delicate balance to articulate the right level of testing
because the "right" level can vary based on each engineer's risk
assessment of a particular change.
Jeff
