[Mlir-commits] [mlir] [MLIR][NVVM] Add support for converting fp4/6/8 to fp16x2 (PR #162439)

[Srinivasa Ravi]

================
@@ -1855,6 +1855,111 @@ def NVVM_ConvertBF16x2ToF8x2Op : NVVM_Op<"convert.bf16x2.to.f8x2"> {
   }];
 }
 
+class NVVM_ConvertF8x2ToFP16x2Op_Base <string dstType>
+: NVVM_Op<"convert.f8x2.to." # !tolower(dstType) # "x2"> {
+  let summary = "Convert a pair of f8 inputs to " # !tolower(dstType) # "x2";
+  let description = [{
+    This Op converts the given f8 inputs in a i8x2 vector to }] # !tolower(dstType) # [{.
+
+    The result `dst` is represented as a vector of }] # !tolower(dstType) # [{ elements.
+
+    The `relu` attribute, when set, lowers to the '.relu' variant of
+    the cvt instruction.
+    
+    [For more information, see PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cvt)
+  }];
+  let results = (outs VectorOfLengthAndType<[2], [!cast<Type>(dstType)]>:$dst);
+  let arguments = !if(!eq(dstType, "F16"),
+    (ins VectorOfLengthAndType<[2], [I8]>:$src,
+         DefaultValuedAttr<BoolAttr, "false">:$relu,
+         TypeAttr:$srcType),
----------------
Wolfram70 wrote:

In this case, this `TypeAttr` is for the source type since it can be either `e4m3` or `e5m2`.

https://github.com/llvm/llvm-project/pull/162439