[Mlir-commits] [mlir] [MLIR] Supported sparse MMA intrinsics in the MLIR->NVVM IR->NVPTX flow (PR #168686)

[Kirill Vedernikov]

================
@@ -2187,6 +2218,29 @@ def MMAIntOverflowAttr : EnumAttr<NVVM_Dialect, MMAIntOverflow, "mma_int_overflo
   let assemblyFormat = "`<` $value `>`";
 }
 
+/// Sparse MMA metadata types
+def MMASpMetadataStandard : I32EnumAttrCase<"standard", 0>;
+def MMASpMetadataOrdered : I32EnumAttrCase<"ordered", 1>;
+def MMASpMetadata : I32EnumAttr<"MMASpMetadata", "Sparse MMA metadata ordering",
+  [MMASpMetadataStandard, MMASpMetadataOrdered]> {
+  let genSpecializedAttr = 0;
+  let cppNamespace = "::mlir::NVVM";
+}
+def MMASpMetadataAttr : EnumAttr<NVVM_Dialect, MMASpMetadata, "mma_sp_metadata"> {
----------------
kvederni wrote:

It has been fixed in 156919a.

https://github.com/llvm/llvm-project/pull/168686