[BUG] [QST] Regression - why Sm90RowBroadcast in 3.5.1 stops support smem usage?

[ankutalev]

Hello!
The question is in the title, proofs

3.5.0
3.5.1

I'm interested in this change, because it breaks backward compatibility and seems suspicious.
Can you provide some details about this change?

Thanks!

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[ankutalev]

This issue still needs to be addressed

[thakkarV]

@jackkosaian @hwu36 CC

[hwu36]

Also CC @richardmcai

[richardmcai]

The static assert comment is inaccurate, if you look at the implementation you see that it does use shared memory. Stage count is still required to be 0 since we no longer pipeline the row vector load across epilogue iterations and instead load it all at once into smem.