[Mlir-commits] [mlir] [MLIR][NVVM] Add support for f8x2 conversion (PR #137781)

[Srinivasa Ravi]

================
@@ -1120,6 +1120,110 @@ def NVVM_CvtToF6x2Op : NVVM_Op<"cvt.to.f6x2"> {
   }];
 }
 
+def CVTFP8E4M3 : I32EnumAttrCase<"E4M3", 0, "e4m3">;
+def CVTFP8E5M2 : I32EnumAttrCase<"E5M2", 1, "e5m2">;
+def CVTFP8UE8M0 : I32EnumAttrCase<"UE8M0", 2, "ue8m0">;
+
+def CVTFP8Type : I32EnumAttr<"CVTFP8Type", "NVVM CVTFP8Type kind",
+  [CVTFP8E4M3, CVTFP8E5M2, CVTFP8UE8M0]> {
+  let genSpecializedAttr = 0;
+  let cppNamespace = "::mlir::NVVM";
+}
+def CVTFP8TypeAttr : EnumAttr<NVVM_Dialect, CVTFP8Type, "cvt_fp8_type"> {
+  let assemblyFormat = "`<` $value `>`";
+}
+
+def NVVM_CvtToF8x2Op : NVVM_Op<"cvt.to.f8x2"> {
+  let summary = "Convert a pair of f32 or fp16 inputs to f8x2";
+  let description = [{
+    This Op converts each of the given float input types to the specified f8 
+    type.
+    The result `dst` is either represented as an i16 type or a vector
+    of two f8 types.
+    The following table describes the supported conversions and their formats:
+    ```
+    |-----------|-----------|--------------------------------------------------|
+    | Src Type  | Dst Type  | Description                                      |
+    |-----------|-----------|--------------------------------------------------|
+    |   f16x2   |  e4m3x2   |  Only operand `a` must be provided and it must   |
+    |           |  e5m2x2   |  be a vector of two F16s.                        |
+    |           |           |  If `dst` is returned as an i16 type, the        |
+    |           |           |  converted values are packed such that the       |
+    |           |           |  value converted from the first element of `a`   |
+    |           |           |  is stored in the upper 8 bits of `dst` and the  |
+    |           |           |  value converted from the second element of `a`  |
+    |           |           |  is stored in the lower 8 bits of `dst`.         |
+    |           |           |  If `dst` is returned as a vector type, each     |
+    |           |           |  converted value from `a` is stored as an i8     |
+    |           |           |  element in the vector.                          |
+    |-----------|-----------|--------------------------------------------------|
+    |   bf16x2  |  ue8m0x2  |  Only operand `a` must be provided and it must   |
+    |           |           |  be a vector of two BF16s.                       |
+    |           |           |  If `dst` is returned as an i16 type, the        |
+    |           |           |  converted values are packed such that the       |
+    |           |           |  value converted from the first element of `a`   |
+    |           |           |  is stored in the upper 8 bits of `dst` and the  |
+    |           |           |  value converted from the second element of `a`  |
+    |           |           |  is stored in the lower 8 bits of `dst`.         |
+    |           |           |  If `dst` is returned as a vector type, each     |
+    |           |           |  converted value from `a` is stored as an i8     |
+    |           |           |  element in the vector.                          |
+    |-----------|-----------|--------------------------------------------------|
+    |  f32, f32 |  e4m3x2   |  Both operands `a` and `b` must be provided and  |
+    |           |  e5m2x2   |  they must be F32 values.                        |
+    |           |  ue8m0x2  |  If `dst` is returned as an i16 type, the        |
+    |           |           |  converted values are packed such that the       |
+    |           |           |  value converted from `a` is stored in the       |
+    |           |           |  upper 8 bits of `dst` and the value converted   |
+    |           |           |  from `b` is stored in the lower 8 bits of       |
+    |           |           |  `dst`. If `dst` is returned as a vector type,   |
+    |           |           |  each converted value is stored as an i8         |
+    |           |           |  element in the vector.                          |
+    |-----------|-----------|--------------------------------------------------|
----------------
Wolfram70 wrote:

That makes a lot of sense. I also think with the different possible operand types currently, it makes more sense for them to be different Ops instead. I will work on this, thanks!

https://github.com/llvm/llvm-project/pull/137781