[PATCH v2] Support Intel AVX10.2 BF16 instructions

[Jiang, Haochen]

> From: Jan Beulich <jbeulich@suse.com>
> Sent: Thursday, October 31, 2024 4:25 PM
> 
> On 31.10.2024 08:54, Jiang, Haochen wrote:
> >
> 
> > The discussion on whether to add NE is still ongoing.
> 
> I see.

The change supposed to be made on mnemonics for BF16 related insts:

VSCALEFPBF16 -> VSCALEFNEPBF16:
Avoid confusion from legacy PS/PD/PH part.

VMINMAXNEPBF16 -> VMINMAXPBF16:
The calculation is always exact and no need for NE.

VCVTTNEBF162I[,U]BS -> VCVTTBF162I[,U]BS:
It is always round toward zero, not an RNE.

> > For VRNDSCALENEPBF16, the NE should be kept after I read the SDM. The
> > ...
> Considering this somewhat peculiar difference I further wonder whether the
> assembler wouldn't better be made warn if the low 4 bits weren't all clear.

That is actually an interesting topic. I am both ok whether to add or not.
It might be an annoying or great reminder for different people.

> Plus actually make explicit what "NE" stands for.

I have raised that since I was confused by this once in the past. They should add
that when doc get revised. I will keep an eye on that.

Thx,
Haochen