[Mlir-commits] [mlir] [AMD][ROCDL][AMDGPU] Support packed conversions fp8/bf8->bf16 and fp8/bf8->fp32 (PR #131850)

[Yi Qian]

yiqian1 wrote:

> To rephrase the portability question - what prevents us from generating the ->bf16 instructions on gfx942?

There are no ->bf16 instructions on gfx942. They're only available on gfx950.

https://github.com/llvm/llvm-project/pull/131850