[Mlir-commits] [mlir] [MLIR][Vector] Add support for distributing masked writes (PR #71482)

[Quinn Dawkins]

qedawkins wrote:

> Hmm, I'm not sure actually this is correct. In my mental model about the masking, it's applied _close_ to the memory, immediately before/after write/read, because that's where masking really happens. So for write, the masking should match with the memory shape, not exactly the vector shape, when we have non-idenity permutation maps. I think we need inverse the map and compose with the vector shape to get the mask shape, and distribute that.. @dcaballe WDYT?

Sorry, I reread the docs for transfer_write and you're right. Moving to a draft for now.

https://github.com/llvm/llvm-project/pull/71482