[Mlir-commits] [mlir] [mlir][AMDGPU] Add scaled floating point conversion ops fp8 (PR #141554)

[Tim Gymnich]

================
@@ -112,6 +112,38 @@ def AMDGPU_ExtPackedFp8Op :
   }];
 }
 
+def AMDGPU_ScaledExtPackedOp
+    : AMDGPU_Op<"scaled_ext_packed", [Pure]>,
+      Arguments<(
+          ins AnyTypeOf<[VectorOfLengthAndType<[2, 3, 4], [F8E5M2, F8E4M3FN]>,
----------------
tgymnich wrote:

> ... and can you point me at the bf16 asymmetry?

There is `cvt.scalef32.f16.fp8` but I could not find `cvt.scalef32.bf16.fp8`.

> (Also, if there are non-packing instructions, we may want to only take even-sized vectors and let -arith-to-amdgpu handle the splitting into packed + non-packed instructions)

According to the function signature `cvt.scalef32.f16.fp8` still seems to be a packing instruction even though its missing the pk in the name.

https://github.com/llvm/llvm-project/pull/141554