> From: Christian Ludloff <[email protected]> > Sent: Wednesday, November 11, 2026 12:23 PM > > On Tue, Nov 10, 2026 at 7:31 PM Jiang, Haochen <[email protected]> > wrote: > > > From: Christian Ludloff <[email protected]> > > > Sent: Wednesday, November 11, 2026 11:10 AM > > > > > > A minor general observation. > > > > > > TOP2PHPS goes best with a TOP2FP16PS alias, for AMX symmetry. (I plan > > > to repeat the comment for nasm and binutils.) > > > > > > https://www.sandpile.org/x86/opc_3.htm#:~:text=TOP2FP16PS > > > > > > And no, I'm not advocating for a general FP16/PH rename – just for > > > this one specific alias, since this case sticks out most. :) > > > > I get your point for legacy AMX. But legacy AMX is the only ISA using > > fp16 for fp16 in some (even not all) mnemonics, all the other ISAs are > > using ph for fp16. > > legacy eh? we'll re-check in, say, a decade? :) > > as for naming... > > camp X likes PH/PS/PD and SH/SS/SD etc. > > camp Y likes FP16/FP32/FP64 and BF16 and all the recent 8/6/4 etc.
The camping understanding here is not fair here. The D for double float, S for single float, H for half float has been used since the first day. We never used FP64/32 in naming, no FP16 naming except for AMX-FP16. That should always be followed. At that time, you could not anticipate how math types will evolve so using letters are fine. Then now we have more math types emerging quickly, and we even have two types on x-bit float. Letters are not straight-forward anymore. It is why numbers started to kick in since BF16. It is caused by different era. But the old naming should still be followed. > > arguably BF16 was the original sin > > then we had VBCSTNEBF162PS, which should have been SBF16, to match > VBCSTNESH2PS I agree it should be SBF16, but it is hard to change after realizing that when we looked back at AVX10.2 (and the NE should also be dropped here). At that time BF16 is still not mature and some mnemonics are failing to match conventions. At least we should minimize those from now on. > > then AMX-FP16's TDPFP16PS, mirroring AMX-BF16's TDPBF16PS As said, they should not be mirrored. It should be TDPPHPS instead in an ideal world. > > and now TOP2xxxPS, at last – luckily the only case sticking out much > So that is why I suppose PH is the right thing here and FP16 should not be an alias. Thx, Haochen > it is what is, sigh :) > > > For me it is AMX not following the consistency, should we really have > > an alias for the inconsistent one? > > and for others... it is ACEv2 not following... should we... > > in the end, where tools are heading now... human preference just > becomes a layer of abstraction – hence the "alias" solution > > under the hood, it's just EVEX F2 5C etc. :) > > -- > C.
