[Mlir-commits] [mlir] [mlir] [XeGPU] Add XeGPU workgroup to subgroup pass (PR #139477)

Adam Siemieniuk llvmlistbot at llvm.org
Mon May 19 05:23:06 PDT 2025


adam-smnk wrote:

> Yes, you are right, the current design expects that the IR uses all the xegpu ops and not any one op in isolation and hence the other ops just follow. This is because these are the cases we have come across so far. > But I agree with you that we might hit the cases you mentioned and I will decouple the distribution logic in subsequent PR from create_nd op for all the ops that operate on vectors .

Fair enough. Could you document it somewhere or at least add TODOs? Or even add a (failing) test case for a standalone `dpas` as it does not behave as expected without `create_nd`.

> I think for ops like load/store/update/prefetch we can safely assume there is a corresponding create_nd.

I'm not sure how that will behave when there are layout conversions. This assumption shouldn't be needed as all ops have necessary information for layout distribution AFAIK.
Anyway, we can iterate on these cases a bit later.



https://github.com/llvm/llvm-project/pull/139477


More information about the Mlir-commits mailing list