[Mlir-commits] [mlir] [mlir][amdgpu] Promote gpu.shuffle to amdgpu.dpp (PR #155158)

[Tim Gymnich]

tgymnich wrote:

To fix the existing implementation and make it consistent with the `gpu.shuffle` spec, we could either:
1. disallow widths that are not equal to the wave size.
2. set `EXEC` according to `width` and return invalid for thread ids >= `width`.
3. just return invalid for thread ids >= `width`.


https://github.com/llvm/llvm-project/pull/155158