WIP: Instruction set detection/dispatch

[drbenmorgan]

This is a small WIP on tools/examples on coding/packaging of instruction set specific code (SIMD etc). At present, it simply implements:

1. A shell script to query the system (macOS/Linux only at present) and print SIMD instructions supported by host.
2. A small C++14 program to do similar

I'm requesting an initial review now to solicit comments of the remaining items:

• How to format SIMD flags for portability, e.g. macOS sysctl gives "SSE4.1", Linux /proc/cpuinfo gives "sse4_1"?
• Demonstration of runtime dispatch/fat binaries. I think this is useful, but also needs documentation on performance penalties, and some actual benchmarks.

Let me know what you think.

[drbenmorgan]

@amadio I couldn't add you as a reviewer, but your feedback would be very welcome here in light of the overlap with VecCore!

[amadio]

Hi @drbenmorgan, interesting project. However, I don't understand the objective that well. Do you want to query SIMD properties of a machine to add proper build flags in the build system? Or do you want to have some way for testing at runtime what is supported to call the right code? I will go through the code with more time and add specific comments later.

For your reference, I gave a talk for the vectorization working group of the IXPUG a while ago, and you can check out the slides here. The IXPUG has lots of resources for this sort of thing. There is also another project made by a Gentoo dev that does part of what you are doing here. It's meant to detect what SIMD is supported by the CPU, so you can add the proper configuration to Portage. It currently supports Intel and ARM CPUs. I think the way it's implemented there is simpler than what is in VCL.

[drbenmorgan]

Hi @amadio,

Hi @drbenmorgan, interesting project. However, I don't understand the objective that well. Do you want to query SIMD properties of a machine to add proper build flags in the build system? Or do you want to have some way for testing at runtime what is supported to call the right code? I will go through the code with more time and add specific comments later.

It's the later more than the former. Given that we'd like to distribute binary packages and these may run on a range of CPU families, what techniques are available to ensure the "compatible and most performant" code is run on a client CPU.

For your reference, I gave a talk for the vectorization working group of the IXPUG a while ago, and you can check out the slides here. The IXPUG has lots of resources for this sort of thing. There is also another project made by a Gentoo dev that does part of what you are doing here. It's meant to detect what SIMD is supported by the CPU, so you can add the proper configuration to Portage. It currently supports Intel and ARM CPUs. I think the way it's implemented there is simpler than what is in VCL.

Thanks, those are very useful! I think this PR as it stands though is more focussed on runtime than build time, and the later could be addressed separately (indeed, part of the project would be to not be smart about selecting flags!).

[amadio]

If your intent is to do runtime checks for SIMD features, I think that implementing something like the intrinsic _may_i_use_cpu_feature from ICC in a way that works for all compilers would be the best way to go. Also, we should map the CPU features to the ones used there. I think that in some places you simplified AVX512 support, which has different versions (e.g. KNL, Skylake) with different subsets supported.

As for selecting flags, if you want a multi-arch binary, you have to select them anyway, so a mechanism needs to be in place for it. Vc has a system to compile for multiple architectures, may be worth having a look.

[drbenmorgan]

@amadio I think I oversold the intent of this PR, so I'll make a few changes to clarify the very limited nature of its aim as a minimal demo (but I agree with your points long term!)