[Mlir-commits] [mlir] [MLIR][XeGPU] Validate DPAS operand types against uArch in XeGPUToXeVM conversion (PR #185081)

[Sang Ik Lee]

silee2 wrote:

> @silee2 before I begin, I just wanted to clarify that we should NOT use the uArch to fix the lowering process (as per this comment [#180107 (comment)](https://github.com/llvm/llvm-project/issues/180107#issuecomment-4009160113)) and instead just focus on the `DpasOp::verify()` function to validate against 32 bit inputs for the lhs and rhs of dpas?
> 
> Are there any other cases we should check for? From the previous conversation in #180107, `i16` is also not supported, but I'm not sure if that is documented somewhere, as it's a valid input type in the TableGen but fails in the conversion lowering process.
> 
> Sorry for all of the trouble, thanks for your patience.

No. Using uArch to check during lowering is fine and actually the recommended way to check. 
The part I asked to revert is not the uArch check itself, but the change related to modifying type conversion related code.
uArch check happens inside the op conversion pattern.
The ask was to revert everything other than that.


https://github.com/llvm/llvm-project/pull/185081