[PATCH 09/10] Support Intel AMX-FP16

[Jiang, Haochen]

> -----Original Message-----
> From: Jan Beulich <jbeulich@suse.com>
> Sent: Wednesday, October 19, 2022 10:13 PM
> To: Jiang, Haochen <haochen.jiang@intel.com>
> Cc: hjl.tools@gmail.com; binutils@sourceware.org; Cui, Lili <lili.cui@intel.com>
> Subject: Re: [PATCH 09/10] Support Intel AMX-FP16
> 
> On 19.10.2022 16:01, Jiang, Haochen wrote:
> >> -----Original Message-----
> >> From: Cui, Lili <lili.cui@intel.com>
> >> Sent: Wednesday, October 19, 2022 6:34 PM
> >>
> >>>>>> -    "CpuAMX_TILE|CpuAMX_INT8|CpuAMX_BF16" },
> >>>>>> +    "CpuAMX_TILE|CpuAMX_INT8|CpuAMX_BF16|CpuAMX_FP16" },
> >>>>>>    { "CPU_ANY_MSRLIST_FLAGS",
> >>>>>>      "CpuMSRLIST" },
> >>>>>> +  { "CPU_ANY_AMX_FP16_FLAGS",
> >>>>>> +    "CpuAMX_FP16" },
> >>>>>>  };
> >>>>>
> >>>>> Same here then.
> >>>> Done.
> >>>
> >>> I guess my comment here was a little misleading (I'm sorry for
> >>> that), in that in addition I was expecting you to consider the
> >>> comment regarding the need for the *_ANY_* constants that I did give
> >>> for several of the
> >> patches in this series.
> >>> I think the question applies here as well: Are there dependent
> >>> features known to appear? If not, there's no need for the extra constant.
> >>>
> >> Get you, there are no known dependencies now, and I will pay
> >> attention on it in future ISAs.
> >
> > But I suppose for AMX-FP16, the possibility of potential dependency is
> > bigger than some like CMPCCXADD, MSRLIST. Since FP16 type is not that rare.
> 
> If you go look you'll find that we don't have ANY_* for a fair part (if not a
> majority) of the features). Plus AMX-FP16 as a feature name suggests more
> generality than there really is - the feature is about a specific multiplication insn,
> not (like e.g. AVX512-FP16) about the introduction of a set of insns on a basic
> data type (whether the set of insns could of course widen down the road, and
> where the further ISAs would depend on the more basic operations on that data
> type).
>

I have your point and it might be more reasonable to predict nothing depends on
that ISA.

Haochen
 
> Jan