Query AVX2 and AVX512VL support when selecting x86 kernels

[martin-frbg]

For #1947 (and corresponding earlier cases involving Haswell) where VM or OS limitations made otherwise standard features of the correctly detected CPU model unavailable.

[yuyichao]

I believe the test here are testing hardware support. (from CPUID).

OS support bit are in XCR0, same as AVX. See the bit assignment in https://en.wikipedia.org/wiki/Control_register#XCR0. In julia (and I believe LLVM) we conservatively check all bits from 5 to 7 (i.e. 0xe0). You can reduce that if you don't need all features I assume....

[martin-frbg]

Indeed, thanks. Hope I got it right now, as I do not have any AVX512 hardware to test this on.

[yuyichao]

Didn't check all the bits but it looks right to me as well....

[brianborchers]

Indeed, thanks. Hope I got it right now, as I do not have any AVX512 hardware to test this on.

I do have a Xeon Skylake system with AVX512 running Linux that I could test on. However, it's not clear that I could provide a complete test, since the main problem here is with OS's that disallow the use of AVX512 instructions on hardware that is otherwise capable.

On my machine, the DYNAMIC_ARCH build should figure out that it has AVX512 and use those kernels. How could I test that this is working correctly?

[mk]

@brianborchers the original issue occurred for me using docker which seems to disable avx512.

I can test this when I’m back at my iMac next week.

[martin-frbg]

If it does not detect a specific capability, the DYNAMIC_ARCH build is supposed to issue a warning about falling back to a less capable cpu. Unfortunately i seem to have missed this in the SkylakeX case, although I preprared the message. Will update this PR in a minute.
Similarly a quick build test should show that the modified getarch writes both HAVE_AVX2=1 and HAVE_AVX512VL=1 to Makefile.conf.

[martin-frbg]

Travis CI seems to be timing out while still installing dependencies on OSX...