[Mlir-commits] [mlir] [mlir][arith] Expand the scaling ops with the value of the scale (PR #217892)

[Krzysztof Drewniak]

krzysz00 wrote:

Re the `cvt.scalef32` - that reads the exponent *if we're dealing with an f8E8M0*. I'd need to double-check if those instructions support E5M3 scales ... but if I remember right, they'll just read the scale bytes out of bits 31:23 of the input, whether or not that'd make sense as a float

https://github.com/llvm/llvm-project/pull/217892