yes, of course,   Re-add the missing mailing list...


在 2023/5/27 13:45, Martin Storsjö 写道:
> Hi,
>
> Thanks - can you send the new patches to the mailing list too? They 
> need to be available publicly for review before they can be accepted. 
> (I didn't check these yet.)
>
> // Martin
>
> On Sat, 27 May 2023, myais wrote:
>
>> Hi, Martin,
>>
>> Thank you for your correction, and I completed the modification 
>> according to your opinion, the attachments are the new patches.
>>
>>
>> Thanks.
>>
>>
>> 在 2023/5/24 20:49, Martin Storsjö 写道:
>>> Hi,
>>>
>>> On Tue, 23 May 2023, myais wrote:
>>>
>>>> Do you have any new opinions here? I am looking forward to your reply.
>>>
>>> I've started looking at this now after focusing on a different issue 
>>> first.
>>>
>>> The big thing is that this is the first new optional instruction set 
>>> on top of aarch64, so there's a bit of work to do to handle that 
>>> properly (with runtime detection, and assembling that code even if 
>>> the baseline target doesn't support it). I've started looking into 
>>> that now.
>>>
>>> In your case your patches don't care about that and just hardcode 
>>> enabling it if the compiler baseline support the instruction, and 
>>> skips it otherwise. I guess that's possibly fine, but your condition 
>>> for the code is wrong; the "usdot" instruction requires the "i8mm" 
>>> extension, not "dotprod". i8mm is part of armv8.6-a (and is 
>>> available on graviton 3, luckily, which allows me to test it).
>>>
>>> So instead of __ARM_FEATURE_DOTPROD, this should use 
>>> __ARM_FEATURE_MATMUL_INT8, and the functions should probably use 
>>> i8mm as suffix instead of dotprod. I guess you can resubmit them 
>>> with that change (and make sure you don't end up with the "no 
>>> trailing newline at the end of file" issue in the changed files in 
>>> any intermediate commit).
>>>
>>>> In addition, I have some new similar patches, which are the aarch64 
>>>> implementations of some other functions, should I wait for your 
>>>> feedback before submitting or submit it directly?
>>>
>>> I'd prefer to settle these patches first before taking on the next set.
>>>
>>> // Martin
>>