[Mlir-commits] [mlir] [MLIR][ROCDL] Add conversion for gpu.subgroup_id to ROCDL (PR #136405)

[Krzysztof Drewniak]

================
@@ -608,6 +608,10 @@ def ConvertGpuOpsToROCDLOps : Pass<"convert-gpu-to-rocdl", "gpu::GPUModuleOp"> {
                clEnumValN(::mlir::gpu::amd::Runtime::HIP, "HIP", "HIP"),
                clEnumValN(::mlir::gpu::amd::Runtime::OpenCL, "OpenCL",
                           "OpenCL"))}]>,
+    Option<"subgroupSize", "subgroup-size", "unsigned",
+           "0",
+           "specify subgroup size for the kernel, if left empty, the default "
+           "value will be decided by the target chipset.">,
----------------
krzysz00 wrote:

Is this a sensible thing to do? You have choosers that can be +wavesize64 and it's quite possible someone might end up not setting this flag. Might be worth seeing if LLVM has something for subgroup size

https://github.com/llvm/llvm-project/pull/136405