VNNI Instructions as AVX2 Instructions #8675
Conversation
|
One thing that I think this PR is still missing is adding a new test that would have failed in the current main branch, but I didn't find an avenue for that yet. |
|
Maybe something in test/correctness/simd_op_check_x86.cpp ? |
|
Thanks for the great comments. I added the test (failing under main) and realized I need to broaden the PR's scope a bit. |
|
@abadams: Thanks for requesting a review from the buildbots. Sorry for asking, but could you provide some guidance on what I should do next? I guess trying to figure out why some tests fail? I'm not yet really familiar with the CI here. |
|
It looks like simd_op_check_x86 failed: https://buildbot.halide-lang.org/master/#/builders/145/builds/167/steps/12/logs/correctness_simd_op_check_x86 I don't see a clear error message though. See if you can reproduce that locally. If you can't, let me know and I'll take a look. |
|
Thanks a lot for the comment! I reproduced the error on an old Broadwell chip. It's an illegal instruction error, so the vnni extension is falsely applied. I could, however, not fix the error yet. I'm a bit confused because AVX512 tests pass without possible hardware support. Do you have a guess for the error's root cause? Otherwise, I'll continue to search anyway. |
|
That test only actually runs code if it detects that it can on the current host CPU. Otherwise it just checks that it compiles successfully. So maybe something is wrong with the host feature detection routine? |
|
Thanks for the feedback, @abadams ! I have updated the code in the test's base class. The test now runs on the Broadwell machine I used and it seems the CI test passes too. Some test still fail though. I wonder if some failures are spurious, but I also wonder if my intervention caused problems for https://buildbot.halide-lang.org/master/#/builders/107/builds/218 . |
|
It's possible that's a real error. The notable thing about the windows bot the failures are on is that it's our only Zen 4 bot. I'm not sure Zen 4 supports avx-vnni. |
Phoronix says Zen 4 is supposed to support |
|
avx512-vnni is a different extension to avx-vnni (and hilariously, avx-vnni came later) |
|
Thanks for the helpful comments! I'll try to reproduce that error. I've two questions in that regard. First, how do you identify the CPU arch of a runner? I looked at the run's output, but couldn't infer that. Second, was there also a Zen4 Linux run? If not, I can try to get hold of a Zen4 Linux machine, hoping the failure is arch and not OS specific. |
Endlessly confusing omg |
src/CodeGen_X86.cpp
Outdated
| t.set_feature(Target::AVXVNNI); | ||
| } | ||
| if (t.has_feature(Target::AVX512_Zen4)) { | ||
| t.set_feature(Target::AVXVNNI); |
Fixes #8673