[Mlir-commits] [mlir] [mlir][xegpu] Convert Vector contraction to XeGPU (PR #122115)

[Adam Siemieniuk]

adam-smnk wrote:

> For downstream, we currently have an arch-aware blocking pass for xetile before lowering xetile to xegpu, and the lowering pass has an arch-specific target to verify the generated xegpu code is valid for that arch. Sounds like you will have the same thing for vector.contract before lowering it into XeGPU?

Correct, my current naive approach is to assume that arch-specific transformations and validation already happened at the vector level. The op conversion is currently treated as a simple direct last mile step shifting responsibility to the user.

It could be better to add some XeGPU utils that help with driving transformations, lowering validation, and even verification of the XeGPU ops. Perhaps, we could start with a hard-coded list of possible combinations (for let's say 2-3 main targets).

https://github.com/llvm/llvm-project/pull/122115