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

[Krzysztof Drewniak]

================
@@ -239,6 +230,65 @@ struct GPUShuffleOpLowering : public ConvertOpToLLVMPattern<gpu::ShuffleOp> {
   }
 };
 
+struct GPUSubgroupIdOpToROCDL final
+    : ConvertOpToLLVMPattern<gpu::SubgroupIdOp> {
+  using ConvertOpToLLVMPattern::ConvertOpToLLVMPattern;
+
+  LogicalResult
----------------
krzysz00 wrote:

I'd prefer that this not be a rewrite pattern on `rocdl`, but a rewrite pattern that's gpu => gpu, and that can be applied before lowering to rocdl

(That way, downstream, we can run this pattern before PRopagateDispatchSizeBounds)

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