[Mlir-commits] [mlir] [mlir][arith] Add LUT-based lowering of arith.extf for FP8 to F32. (PR #196321)

[Krzysztof Drewniak]

https://github.com/krzysz00 commented:

Various nitpicks + "hey, we should try this in IREE once it lands" + "can you do FP4 while you're here?"

Also, you can support vectors by unrolling to scalars ... though, honestly, if you've got 1-D vector, all you need to do is replace the `memref.load` with a `vector.gather` and I'd consider doing that instead. (N-D vectors you could probably shape_cast to 1-D)

https://github.com/llvm/llvm-project/pull/196321