[Mlir-commits] [mlir] [MLIR] Add f8E8M0FNU type (PR #111028)

[Jakub Kuderski]

kuhar wrote:

> I would say they are still floating-point in spirit :). I’m not sure we could treat them differently. There are specific FP semantics that still apply to these formats, such as NaN propagation (e.g., minnum/maxnum/minimum/maximum), exceptions, overflow...

Is there any hardware that supports arithmetic over these types? At least on the *CDNA* side, fp8 support is limited to conversion to/from other 'usual' fp types and dedicated matmul intrinsics. If we had IR that did any arithmetic on fp types, we wouldn't be able to compile it without software emulation anyway. Or in other words, I wonder if some level of *general* hw support (not just conversions + matmul) would be a good criteria for inclusion in core MLIR types.

https://github.com/llvm/llvm-project/pull/111028