[QST] Row Contiguous C Matrix from TiledMMA?

[osayamenja]

What is your question?
Hello!

How do we get a row contiguous value layout from TiledMMA? That is, referring to this LaTeX value layout, we can see that values are interspersed across threads(I believe this helps coalescing?) How can we permute that value layout to achieve column-major contiguity, where Thread 0 owns Vals 0 to N-1; Thread 1, Vals N to 2N - 1 and so on?

My use case requires an independent epilogue operation performed on each row of the C matrix and, naturally, we can parallelize across threads. Currently, I permute the value layout by copying from registers to a shared memory workspace. However, this copy would be unnecessary, if the C matrix values were thread-row contiguous. In addition, I would also be able to operate on the register fragment directly rather than shared memory.

There probably is a memory access tradeoff lurking in the answer to this question, so please let me know if that is the case too :)

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.