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

Krzysztof Drewniak llvmlistbot at llvm.org
Thu Jun 5 11:17:48 PDT 2025


================
@@ -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]>,
----------------
krzysz00 wrote:

I'm more thinking that if there _isn't_ a scalar instruction, we'll want to pad with 0s, but if there is one then never mind.

... and can you point me at the bf16 asymmetry? I can't think of one off the top of my head

(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)

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


More information about the Mlir-commits mailing list