Add AVX10.2 Intrinsics

[sayantn]

Tracking Issue: rust-lang/rust#153417

Note: requires rust-lang/rust#153068, contains a temporary fix in order to pass tests

• zero-extending partial moves
• sum of absolute differences
• dot product
• saturating truncating float->integer conversions
• minmax
• saturating float->int8 conversions
• packing float32->float16 conversions
• float16->float8 conversions with bias
• float16->float8 conversions
• float8->float16 conversions
• SM4 intrinsics
• Read-shared move intrinsics
• bfloat16 instructions
  • bfloat16 minmax
  • saturating bfloat16->int8 conversions

BFloat16 intrinsics cannot be done now, because the LLVM intrinsics use bf16 vectors, and we cannot link to them directly. Requires rust-lang/rust#140763 or some other workaround.

AVX10_V1_AUX intrinsics

These instructions are supported in a subset of AVX10.2, but neither GCC nor LLVM supports this feature yet.
TODO: Modify the #[target_feature] attribute once LLVM starts supporting this.

• integer dot product instructions
• packing float32->float16 conversions
• float16->float8 conversions with bias
• float16->float8 conversions
• float8->float16 conversions

LLVM Bugs

• single-element minmax instructions: Miscompilation of VMINMAXS{H,S,D} intrinsics llvm/llvm-project#184245
• saturating float->int8 conversions: Miscompilation of VCVT{T}P{H,S}2I{U}BS intrinsics llvm/llvm-project#184251
• hf8->float16 conversions: Miscompilation of cvthf8_ps intrinsics llvm/llvm-project#184651

[sayantn]

@Amanieu @folkertdev what do you think, can/should this be merged now, considering the LLVM bugs, or should we wait until the fixes merge?

[folkertdev]

I don't have strong feelings about it. I'm assuming this won't really see practical usage? On the other hand we might be adding some test coverage to potentially catch regressions next LLVM upgrade.

So I guess I'm fine with merging, but I don't think it matters much really.