[C++] Use existing library for CPU detection

[pitrou]

Describe the enhancement requested

We currently use our own home-grown module for detecting CPU features, but it mainly supports x86, which implies that for now we cannot do dynamic dispatch on non-x86 CPUs (for example ARM where we might want to detect ARM or even SVE support at runtime).

We could probably reuse https://github.com/google/cpu_features instead. It doesn't seem to have any dependencies, and could be bundled as a private static library.

This was already proposed years ago as #16715

Component(s)

C++

[pitrou]

Looking at https://github.com/google/cpu_features/blob/main/include/cpuinfo_aarch64.h , it's not obvious to me how we could get the supported SVE vector widths. Do you have any idea @cyb70289 ?

[pitrou]

Edit: apparently there's an intrinsic to get the vector width: https://lemire.me/blog/2022/11/29/how-big-are-your-sve-registers-aws-graviton/

[kou]

+1

[cyb70289]

+1

[wgtmac]

The library itself looks good. My main concern is the slowdown of its release cadence: https://github.com/google/cpu_features/releases

[pitrou]

Hmm, that's a good point. Maintenance also seems to have slowed down a bit: https://github.com/google/cpu_features/graphs/contributors?from=20%2F01%2F2024

[pitrou]

There's also https://github.com/anrieff/libcpuid , which only supports x86 and ARM but exposes a lot of information.

[pitrou]

And by the way, xsimd has a more limited functionality for CPU feature detection: https://xsimd.readthedocs.io/en/latest/api/arch.html . It only supports x86 and ARM, does not offer detailed CPU information, does not expose things like BMI2 etc.

[h-vetinari]

For completeness, you might be interested in https://github.com/archspec/archspec - it supports x64, aarch, ppc & risc, plus a lot of meta-information about the (partial) inclusion ordering of the various CPUs in terms of feature sets.

I think it doesn't have a C/C++ API though, so this is probably a long shot (slow to call python from C++), but OTOH, querying CPU features is not a common occurrence (once per session...?) and could be cached.

Numpy does it the other way around, detecting individual CPU features, rather than processor models, with a result à la:

Supported SIMD extensions in this NumPy install:
    baseline = SSE,SSE2,SSE3
    found = SSSE3,SSE41,POPCNT,SSE42,AVX,F16C,FMA3,AVX2
    not found = AVX512F,AVX512CD,AVX512_KNL,AVX512_KNM,AVX512_SKX,AVX512_CLX,AVX512_CNL,AVX512_ICL

It's implemented in C, but I don't think it has a public API.

[pitrou]

@h-vetinari Well, we don't want the core Arrow C++ library to depend on the existence of a Python interpreter and Python libraries. As for the NumPy C code, since it's internal it's probably not trivial to vendor, and I suppose its APIs have no stability guarantee either.

[h-vetinari]

Yeah, I understand. That's why I said "for completeness". :)

[pitrou]

The library itself looks good. My main concern is the slowdown of its release cadence: https://github.com/google/cpu_features/releases

For the record, a new release was just made (v0.10.0).

[pitrou]

Also cc @AntoinePrv