Implement runtime dispatch of SIMD code

[althonos]

Hi @nicodr97 !

Sorry for the small delay in getting this implemented, I finally got to sit down and work on some extra code besides what I need for my PhD 😄

I added the cpu_features from Google as a git submodule, and updated the CMake files to compile it as a subdirectory. cpu_features is available under the Apache 2.0 license so there is no licensing issue here. When a new Manager is creeate, it checks the local CPU feature flags to check what backend is supported, using a mixture of compilation flags and runtime flags:

• on x86, it checks for SSE2 and AVX2
• on x86-64, SSE2 is always available, it checks for AVX2
• on arm, it checks for NEON
• on aarch64, NEON is always supported.

It's still possible to disable AVX2, NEON, or SSE2 at compile time; but now a single binary can be compiled and distributed for each platforms. I also added an [INFO]-level message to report when an optimized implementation of the statistics is being used, to make it easier to debug.

[nicodr97]

Very nice!

Given that the platform value depends on the specific compilation parameter passed by the user, is it necessary to check it all around? That is, doesn't platform == ComputePlatform::SSE2 means that HAVE_SSE2 is defined?

Thank you for your contribution!

[althonos]

Given that the platform value depends on the specific compilation parameter passed by the user, is the necessary to check it all around? That is, doesn't platform == ComputePlatform::SSE2 means that HAVE_SSE2 is defined?

Yes, platform == ComputePlatform::SSE2 means HAVE_SSE2 is defined; however you cannot remove the include guards because the SSE2 statistics (e.g. SSE2Similarity) may not be defined if HAVE_SSE2 is not defined. If you remove the include guards, you'll get compile errors because some classes are not defined.

You could include stub files to still compile a mostly empty SSE2Similarity in have HAVE_SSE2 is not defined, but that's definitely more work than just add a few include guards at the right locations for basically the same result.