coll/xhc: Bring in latest developments #12908
Conversation
|
I assume the mpi4py CI failure here is unrelated to this PR. |
|
No this is due to a known race condition issue unrelated to this pr. See my latest pr against main which addresses the race condition. |
|
I would honestly look at it slightly differently, we can discuss these issues ad absurdum, in my opinion they are not earth shaking either way (that's why I marked them as minor). The same rules are not being used in a number of other projects that I am involved, so I truly do not hold any religious beliefs in them. However, I do see a benefit for a multi-institutional project such as Open MPI to follow some coding guidelines, to avoid that different parts of the code look vastly different. |
|
I agree it's not the end of the world; I did kind of go and open a can of worms here, and I can't reasonably argue against adhering to an agreed upon standard. There are some issues associated with doing these conversions, so I would avoid them if it was agreed that the specific guideline(s) are not as heavily enforced/required/necessary (??), but it can certainly be done. |
|
Guess I'm more aligned with Edgar on this one - not religious about it, but there is a benefit to having coding standards, and it makes little sense to have them if we start saying "well not for this PR". 🤷♂️ Ultimately up to the community. |
|
That's understandable -- I'll do the conversion, thanks for the input |
gkatev
force-pushed
the
xhc_v2
branch
2 times, most recently
from
3f27c83 to
f22f327
Compare
Multiple changes, new features, and enchancements Notable new features: - Inlining data in control path for extra small messages. - Allreduce goes through Bcast's implementation intead of duplicating it. - Fine grained op-wise tuning for hierarchy, chunk size, cico threshold. - Reduce reworked & improved, with double buffering and other optimizations. - When smsc support is not present, handle only small messages (below cico threshold), instead of not at all. Signed-off-by: George Katevenis <[email protected]>
|
I've rectified the coding style issues, this is again ready for review. Did you maybe have any other comments for this PR? @bosilca did you maybe have a chance to look this over? (sorry for the 'explicit ping', it appears I can't 'request a review' here, I can only re-request). |
Multiple changes, new features, and enchancements
Notable new features: