return architecture levels for micromamba

[isuruf]

cc @beckermr, @h-vetinari, @jaimergp, @wolfv

[jonashaag]

Is there a way to unit test this?

[JohanMabille]

Is there a way to unit test this?

We could check preprocessor tokens (like we do in xsimd) to detect the architecture, and check that result of get_archspec_x86_64 is consistent, but that would test only the architecture available on the CI. If we want to be exhaustive, we could explictly pass the arch we want to use in the CI (see xsimd again ;) for an example). However, that would require a lot of new jobs in the CI and I'm not sure we want that for testing a single function given the time the CI is already taking.

[jonashaag]

Maybe we can at least compare the output of the Mamba logic to the output of the xsimd logic? That's still limited to the runner architecture but it's better than nothing

[mbargull]

On Linux you could use QEMU. E.g., if you were to compile a test program that prints this function's output:

[ "$( qemu-x86_64 -cpu Penryn test-get_archspec_x86_64` )" = "x86_64" ]
[ "$( qemu-x86_64 -cpu Nehalem test-get_archspec_x86_64` )" = "x86_64-v2" ]
[ "$( qemu-x86_64 -cpu Haswell test-get_archspec_x86_64` )" = "x86_64-v3" ]
[ "$( qemu-x86_64 -cpu Skylake-Server test-get_archspec_x86_64` )" = "x86_64-v4" ]

This would only work if the host CPU supports the features too. I.e., depending on the runner, you may want to guard the test for Skylake-Server at least.
And this would only test on Linux. No idea what the emulators on macOS offers.

[isuruf]

I think adding a better test would not be worth the trouble. If we use __builtin_cpu_supports in the test, we'll be duplicating the actual code in test code which doesn't really test anything. If we use preprocessor macros like __AVX2__, we need to set -march=native, but we need quite a bit of logic to make that work. (Check that it is UNIX and (clang or gcc) and x86_64 which is not easy in CMake as there are options like CMAKE_OSX_ARCHITECTURES)
Finally using an emulator like QEMU or Intel SDE would work, but not sure it's worth it to complicate the CI infrastructure.

[JohanMabille]

Indeed, I missed that we don't pass --march=native, and other options would make the CI much more complicated for a simple function. Let's merge as is.