[PATCH] x86_64: memcpy/memmove family optimized with AVX512

[Carlos O'Donell]

On 01/13/2016 01:10 PM, Senkevich, Andrew wrote:
>> On 12-01-2016 12:13, Andrew Senkevich wrote:
>>> Hi,
>>> 
>>> here is AVX512 implementations of memcpy, mempcpy, memmove, 
>>> memcpy_chk, mempcpy_chk, memmove_chk. It shows average
>>> improvement more than 30% over AVX versions on KNL hardware,
>>> performance results attached. Ok for trunk?
>> 
>> It is too late for 2.23, but ok after review for 2.24.
> 
> We would like this patch to be considered for glibc 2.23 since the
> functionality completes AVX-512 improvements of mem* routines. Memset
> tuned for AVX-512 is already checked in so it looks reasonable to
> have full support in 2.23. Also the changes are strongly AVX-512
> specific, not adding any new interfaces so potential risk of the
> patch is pretty low.
> 
> We already got review comments without any major questions to the
> patch and fixed version will be ready today.
> 
> Given all this can the patch go to current glibc trunk after review
> is finished?

I've complete my review and the patches look good. I had one question
which should not block acceptance for 2.23 (wanted to know about the
scalability of AVX512 across threads in a process and how the routines
perform versus the others on a scalability perspective e.g. some
functional groups might not scale well due to interlocks and other
reasons).

Given that this work is half complete I think we should accept Intel's
patches and complete the work for 2.23.

A NEWS note is required to highlight the new support for AVX512.

Cheers,
Carlos.