Ben Vanik

There should have been no material changes from that change in a correctly operating backend in legacy_sync mode so I'm not sure looking at IR will help much, but if...

That won't work for long as it's not retaining the semaphore it's returning - but more importantly you can't share semaphores like that as that would cause timelines to go...

This information is available here: https://iree-org.github.io/iree/deployment-configurations/ That could be updated to have a better overview without needing to click in to each backend, though.

I think ideally the mmt4d would map to the output buffer and min would run on it in-place - so there should be no extra allocations required (currently it's allocating...

Hi @sogartar! This is likely to require higher-bandwidth than an issue and I know there are others looking into this as well who may want to participate. I can try...

All of _that_ said, the original design outlined in the issue had the goal of being quick and easy to implement and then delete as we learned more about how...

yeah, this is something that'd have to be fixed in the user program - we're seeing the warning at runtime pointing at the vm ops but the issue is in...

I'm not sure it's vmvx-specific, just that vmvx is the only place we can see this happen with ubsan - if it's a multiply that needs a full 64-bit result...

Wanted to raise our hand as someone who needs blocking operations on workers as well! ✋ (we're compute-focused and don't have concepts of frames or clearly delineated times when we...

An interesting discussion in the Dawn matrix room with @austinEng where he had a really nice idea and I wanted to document here: There may be a single primitive that...