[Valgrind-developers] AVX-512 enabling progress

[Mineeva, Tatyana A]

Hello Julian,

We have implemented a number of AVX-512 instructions in Valgrind core.
NAS IS benchmark, compiled with AVX-512, runs correctly under Nulgrind. Tests for the implemented AVX-512 instructions (similar to the tests in ./none/tests/amd64/avx2-1.c) also work correctly.

For now, neither the test nor the benchmark works under Memcheck, because a lot of new IRs are introduced. And a number of AVX-512 instructions is not implemented yet.

1. What has higher priority, Memcheck correctness or full instruction set coverage?
2. When should we start submitting patches: when everything is added and tested on more benchmarks, or earlier, in order to review a work-in-progress?

Thank you for any advice.
Regards, Tanya

--------------------------------------------------------------------
Joint Stock Company Intel A/O
Registered legal address: Krylatsky Hills Business Park,
17 Krylatskaya Str., Bldg 4, Moscow 121614,
Russian Federation

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.

[Julian Seward]

Hi Tanya,

Sorry for the slow reply. I was OOo last week.

Post by Mineeva, Tatyana A
We have implemented a number of AVX-512 instructions in Valgrind core.
NAS IS benchmark, compiled with AVX-512, runs correctly under Nulgrind.
Tests for the implemented AVX-512 instructions (similar to the tests in
./none/tests/amd64/avx2-1.c) also work correctly.

Good.

Post by Mineeva, Tatyana A
1. What has higher priority, Memcheck correctness or full instruction set coverage?

Instruction set coverage. Getting Memcheck to handle new IRs is much easier
than adding new insn support.

Post by Mineeva, Tatyana A
2. When should we start submitting patches: when everything is added and
tested on more benchmarks, or earlier, in order to review a work-in-
progress?

Earlier is better. Even if they don't get landed immediately, they can be
looked at by others. In particular I'd like to see what new IROps you had
to add.

Do you have a bug report (bugs.kde.org) tracking this, to which you can attach
the patches? That is the usual way of doing reviews.

J