[C++] Building AVX512 code when we should not

[asfimport]

On autobrew/homebrew-core#31, where we are packaging Arrow for an old macOS SDK version, we found what I believe are 2 different problems:

1. The check for AVX512 support was returning true when in fact the compiler did not support it
2. Even when we manually set the runtime SIMD level to less-than-AVX512, it was still trying to compile one of the AVX512 files, which failed. I added a patch that made that file conditional, but there's probably a proper cmake way to tell it not to compile that file at all

cc @cyb70289 @pitrou

Reporter: Neal Richardson / @nealrichardson
Assignee: Frank Du / @jianxind

PRs and other links:

• GitHub Pull Request #8478

_{Note: This issue was originally created as ARROW-10321. Please see the migration documentation for further details.}

[asfimport]

Antoine Pitrou / @pitrou:

The check for AVX512 support was returning true when in fact the compiler did not support it

You should point to the logs so that we can look at the symptoms.

[asfimport]

Neal Richardson / @nealrichardson:
Sure, though all of the build logs and discussion is available on the PR I linked.

[asfimport]

Yibo Cai / @cyb70289:
Maybe similar problem as ARROW-9877?
I didn't find build logs at the autobrew PR link.

[asfimport]

Neal Richardson / @nealrichardson:
The X's next to the commits go to Travis builds that failed.

Here's the first, where it builds with -DARROW_HAVE_RUNTIME_AVX512 when it should not, and it fails: https://travis-ci.org/github/autobrew/homebrew-core/jobs/736104290#L723

Here's the second, where I set the runtime SIMD level to AVX2 so it didn't have that flag, yet it still tried to compile util/bpacking_avx512.cc: https://travis-ci.org/github/autobrew/homebrew-core/jobs/736173702#L725

Then you can see in the discussion on the PR I applied a patch that wrapped that whole file in #if defined(ARROW_HAVE_RUNTIME_AVX512), and the next build compiled successfully

[asfimport]

Neal Richardson / @nealrichardson:
It may be similar to ARROW-9877 but that patch is included in 2.0.0, so that wasn't sufficient to fix this. In this case we're building on an older macOS SDK version than what Homebrew generally supports.

[asfimport]

Frank Du / @jianxind:
For the second issue, yes, it should be better fixed in cmake side, I will proposal a patch soon.

For the first issue, seems like the avx512 flag is supported but not full avx512 intrinsic APIs exist for this compiler. Currently we are simply using check_cxx_compiler_flag for the simd detection, it could be more accurate to use check_cxx_source_compiles to perform a try-build on a example SIMD code.

[asfimport]

Neal Richardson / @nealrichardson:
Issue resolved by pull request 8478
#8478

[asfimport]

Sanchit Jain:
Unrelated to this issue, but why does the current build process use AVX2 as the default SIMD level, even if AVX-512 is supported on a machine?

[asfimport]

Antoine Pitrou / @pitrou:
Which "default SIMD level" are you talking about? The default compile-time level is currently SSE4.2.