We currently emit a warning if you try to use float16/bfloat16 without native support, but (surprise!) people ignore warnings (or they are swallowed by the build system), and then wonder why the code is so slow.

I'd propose that we instead consider removing the warnings, and add a new Feature flags named (say) for "slow_emulate_f16" and "slow_emulate_bf16":

• using f16 or bf16 without a proper enabling flag of some sort is an error, not a warning
• specifying these flags in combination with a native f16 flag (eg arm_fp16, f16c) is an error

This requires explicit user understanding that they are building code that is functional, but not performant.