[PATCH] D135202: [IR] Add a target extension type to LLVM.

Wed Jan 11 20:15:32 PST 2023

zixuan-wu added a subscriber: lattner.
zixuan-wu added a comment.

In D135202#4042268 <https://reviews.llvm.org/D135202#4042268>, @pengfei wrote:

> In D135202#4040194 <https://reviews.llvm.org/D135202#4040194>, @jcranmer-intel wrote:
>
>> In D135202#4038774 <https://reviews.llvm.org/D135202#4038774>, @zixuan-wu wrote:
>>
>>> I'd like to enable the backend infra and propose draft later. To better demonstrate and enable backend target-independent code completely, I'd like to demonstrate with x86_amx that changing x86_amx type to target extension type and make all test passed, if nobody objects.  @jcranmer-intel @LuoYuanke @craig.topper @pengfei
>>
>> I have no objections to this; indeed, my original proposal for target extension types included `x86_mmx` and `x86_amx` as types that could be implemented as target extension types. However, I'm not involved with the x86 backend, so I must defer to the proper codeowners there as to whether or not to do this.
>
> I think it's more appropriate to start with `x86_fp80` and `x86_mmx`. I'd expect there will be a more general matrix representation for the AMX type, e.g., TLX <https://discourse.llvm.org/t/rfc-proposal-for-tlx-tensor-llvm-extensions/59261>, <8x8xi8> <https://discourse.llvm.org/t/intel-amx-vector-dialect/2984> etc. rather than limit it to target extension type, though I'm not against to do prototype for it.

Well, as chris's @lattner comment of TLX <https://discourse.llvm.org/t/rfc-proposal-for-tlx-tensor-llvm-extensions/59261> which I also prefer, it's not a good way to support high level type in LLVM IR such as Matrix(or multi-dimension vector). We can convert matrix to fixed vector type and leverage with those intrinsics matrix-intrinsics <https://llvm.org/docs/LangRef.html#matrix-intrinsics>, if we need fixed sized matrix. I am not sure whether it's enough and workable to completely develop AMX feature. For example, need we distinguish vector type and matrix when we addRegisterClass? And another point is do we really need different shape matrix in LLVM IR or just one type called x86amx?

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135202/new/

https://reviews.llvm.org/D135202