[QST] TMA copy predicate

[hyhieu]

Hi everyone,

I have a conceptual question: in CuTe, is it possible to do predicate copy with TMA ops such as SM90_TMA_LOAD? If yes, what is the approach? Perhaps something like copy_if().

Thanks!

[thakkarV]

TMA itself supports predication. If you pass in a base coordinate whose tile lies outside of the tensor shape in the TMA descriptor, those OOB loads will be predicated off by TMA itself.