Add framework for implementation of support for AVX512VNNI

[stweil]

Using AVX512VNNI instructions is still missing.

[stweil]

I am curious: what is the use of -mavx512vl? Should it be checked in configure, too?

[amitdo]

It does not compile without the second flag. It is needed because the code uses ymm registers (256 bit) instead of zmm registers (512 bit).

Yes, we can also add another check in configure.

Note that AFAIK all Intel's CPUs that support AVX512VNNI also support AVXVL.

[stweil]

Thanks for explaining. Maybe we don't need the additional check then.

[amitdo]

Ready for testing.

I suggest to backup the home directory and any other important stuff, in case this patch will trigger a hard crash (reboot).

[stweil]

I suggest to backup the home directory and any other important stuff, in case this patch will trigger a hard crash (reboot).

Will it? I cannot test it before August 21 because I currently don't have access to the required hardware.

[amitdo]

Will it?

I don't think so...

[amitdo]

Stefan, will you able to test the code today or in the upcoming days?

[stweil]

I just finished a first run of intsimdmatrix_test with the new code. It works, but I was surprised because the performance is identical to the AVX2 implementation.

[stweil]

Support for sw build is still missing. Therefore CI currently fails.

[amitdo]

I was surprised because the performance is identical to the AVX2 implementation.

This is disappointing.

We can also remove the 'ones' vector. The vnni version does not need it.

[amitdo]

Did you fetch and merge my commit locally?

git switch avx512-vnni
git log -1

Do you see my commit?

[stweil]

We can also remove the 'ones' vector. The vnni version does not need it.

I removed it locally. But that did not change the generated code. Obviously the compiler also removed it.