Implement Sapphire Rapids features

[jwlawson]

It would be good to have support for the additional AVX features added with Sapphire Rapids, including VNNI, BF16 and AMX support.

• [x] Add Sapphire Rapids target feature (#5677)
  • [x] Extend CPUID support in src/runtime to allow querying cpuid(eax=7, ecx=1) (#5684, #5702)
• [x] Add support for avx512 bf16 (see Intrinsics guide)
  • [x] Add support for vcvtneps2bf16 (#5677)
  • [x] Provide wrapper for 128bit version of vcvtneps2bf16, as LLVM does not provide an unmasked version of the intrinsic (#5704)
  • [x] Add support for vcvtne2ps2bf16 (#5711)
  • [x] Add support for bf16 dot product (#5712)
• [ ] Add support for avx512 VNNI (see Intrinsics guide)
• [ ] Add support for AMX matrix extensions (see Intrinsics guide)

[dsharletg]

Whoops didn't mean to edit, was just trying the check boxes...

[changhoon-sung]

Sapphire Rapids also supports AVX512-FP16, but I noticed that it’s not included in the detected feature list. Is there a specific reason for this exclusion?

[abadams]

There is no specific reason. It's just not yet implemented. PRs welcome.