> -----Original Message----- > From: Andi Kleen <a...@linux.intel.com> > Sent: Saturday, August 3, 2024 3:06 AM > To: Jiang, Haochen <haochen.ji...@intel.com> > Cc: gcc-patches@gcc.gnu.org; Liu, Hongtao <hongtao....@intel.com>; > ubiz...@gmail.com > Subject: Re: [PATCH 0/1] Initial support for AVX10.2 > > > > > INT8 is actually char per my understanding. > > > > For FP8, currently there is no basic calculation insts yet. So we have no > > support for them in AVX10.2 currently, and treat them just as a piece > > of char. > > > > Also there might be other issues for FP8 to discuss, like ABI issues, so > > we put the support aside for now. When everything is mature, we may > > add the support for that. > > But then it's too late isn't it? You wouldn't be able to change > the types of the existing intrinsics anymore, or later end up with
Hi Andi, For bf16 type adding in x86, we actually did not add another set of intrins for the memory type to real type conversion, but reported warning or maybe note. See https://gcc.gnu.org/gcc-13/changes.html But you are right, it is somehow nasty when doing that change in the future. People need to rewrite the code. BTW, I noticed that in LLVM there is FP8 support for ARM currently undergoing. I will have a look on it to see if everything is mature. Maybe need some more input. Thx, Haochen > two sets of intrinsics, and end up with interoperability problems > with full computation. > > Better to define proper types from the beginning. > > -Andi
