[Mlir-commits] [mlir] [mlir][tosa] Add F32 to F16 type narrowing pass (PR #218900)

[Luke Hutton]

================
@@ -247,11 +255,16 @@ FailureOr<Attribute> convertDenseResourceElementsAttr(
     ShapedType type, DenseResourceElementsAttr attr,
     const TypeConverter &typeConverter, bool allowLossyConversion) {
   static_assert(Kind == TosaNarrowKind::Int64ToInt32 ||
-                Kind == TosaNarrowKind::Float64ToFloat32);
-  using From =
-      std::conditional_t<Kind == TosaNarrowKind::Int64ToInt32, int64_t, double>;
-  using To =
-      std::conditional_t<Kind == TosaNarrowKind::Int64ToInt32, int32_t, float>;
+                Kind == TosaNarrowKind::Float64ToFloat32 ||
+                Kind == TosaNarrowKind::Float32ToFloat16);
+  using From = std::conditional_t<
+      Kind == TosaNarrowKind::Int64ToInt32, int64_t,
+      std::conditional_t<Kind == TosaNarrowKind::Float64ToFloat32, double,
+                         float>>;
+  using To = std::conditional_t<
+      Kind == TosaNarrowKind::Int64ToInt32, int32_t,
+      std::conditional_t<Kind == TosaNarrowKind::Float64ToFloat32, float,
+                         uint16_t>>;
----------------
lhutton1 wrote:

Is it possible to use an APFloat as the underlying type for FP16 as opposed to uint16_t? I'm thinking this will make things slightly more readable

https://github.com/llvm/llvm-project/pull/218900