[Mlir-commits] [mlir] [MLIR][NVVM] Support kind::ti16 for tcgen05.mma Ops (PR #216588)

[Durgadoss R]

================
@@ -5609,6 +5609,26 @@ mlir::NVVM::IDArgPair TensormapReplaceOp::getIntrinsicIDAndArgs(
 // NVVM tcgen05.mma functions
 //===----------------------------------------------------------------------===//
 
+static llvm::nvvm::Tcgen05MMAKind
+getNVVMTcgen05MMAKind(NVVM::Tcgen05MMAKind kind) {
+  switch (kind) {
+  case NVVM::Tcgen05MMAKind::F16:
+    return llvm::nvvm::Tcgen05MMAKind::F16;
+  case NVVM::Tcgen05MMAKind::TF32:
+    return llvm::nvvm::Tcgen05MMAKind::TF32;
+  case NVVM::Tcgen05MMAKind::F8F6F4:
+    return llvm::nvvm::Tcgen05MMAKind::F8F6F4;
+  case NVVM::Tcgen05MMAKind::I8:
+    return llvm::nvvm::Tcgen05MMAKind::I8;
+  case NVVM::Tcgen05MMAKind::TI16:
+    return llvm::nvvm::Tcgen05MMAKind::TI16;
+  case NVVM::Tcgen05MMAKind::MXF8F6F4:
+  case NVVM::Tcgen05MMAKind::MXF4:
+  case NVVM::Tcgen05MMAKind::MXF4NVF4:
----------------
durga4github wrote:

I believe, these kinds are used with "Block-scale" and hence are not supported here.
Can we add a comment on this here?
(I ask because they are all values from the same enum, but some are valid, while others are not.)

https://github.com/llvm/llvm-project/pull/216588