[Mlir-commits] [mlir] [MLIR][NVVM] Add Float to TF32 conversion Op (PR #123199)

[Durgadoss R]

durga4github wrote:

> We should have an NVVM OP for this conversion since we have PTX. LGTM!
> 
> Orthogonal topic: We also need a generic conversion operation that can be used with vectors. This operation should convert anything to anything. We should place that operation in the `NVGPU dialect`, as it's a bridge dialect. We can revisit this problem when we need it. I'm just thinking out loud here.

Yes, totally agree. I have a few more CVT ops to add matching the available intrinsics/PTX.
Once we have a set of them, I believe, we will have a better picture (a clear pattern) to build NVGPU Op(s).

https://github.com/llvm/llvm-project/pull/123199