[Mlir-commits] [mlir] [mlir][VectorToGPU] Register NVGPU only when requested (PR #236134)

[Markus Böck]

https://github.com/zero9178 approved this pull request.

LGTM for the error you're seeing

In an ideal world you'd probably actually make a separate pass (VectorToNVGPU?) such that consumers of VectorToGPU never need to build and link in the NVGPU dialect and any of its dependencies. I'd encourage that as a follow up that has even more gains

https://github.com/llvm/llvm-project/pull/236134