[FINUFFT] Microarchitecture expansion

[ludvigak]

The FINUFFT library doesn't have CPU instruction set dispatch, but benefits a lot from optimization with SIMD instructions, so compiling it for different microarchitectures seems like a good idea.

[giordano]

Note that if you change julia_compat you also must change the version of the package itself, unfortunately the registry won't like having different compatibility bounds when we only change the build number

[ludvigak]

Hmm, but that version is supposed to match the version of the binary library. Or do you suggest that the version of finufft_jll diverge from the version of libfinufft?

[giordano]

I know, but some times we do have to diverge, unfortunately. Or maybe you can try to nicely ping developers to tag a new release, there have been some changes since last release already 🙂

[ahbarnett]

I have just pushed 2.0.4. Do I need to tag a release? I'm waiting on fixing a python wheels CI failure that shouldn't affect anything you're doing.

[giordano]

Lovely, thanks a lot! 😃

[giordano]

We should probably add a way to automatically expand microarchitectures only for specified architectures, it isn't uncommon that the libraries do this kind of optimisation only for x86_64.

If you were worried about the failures on aarch64, those are bugs that we have to fix anyway (I already have some fixes locally but have to test something else)

[ludvigak]

Well yes, I was trying to circumvent the failures :)

But in any case, the goal of the microarchitecture expansion is to get binaries with AVX support, so limiting it to x86_64 makes a lot of sense in any case.

[giordano]

For the record, I'm slowly making some progress towards #4252. JuliaPackaging/BinaryBuilderBase.jl#205 fixed the aarch64 flags (not relevant here, but important before moving microarchitectures forward) and with JuliaPackaging/BinaryBuilderBase.jl#232 it is now possible to restrict the expansion with an additional argument:

julia> platforms_x86_64 = filter(p -> p.tags["arch"]=="x86_64", supported_platforms());
       platforms_other  = filter(p -> p.tags["arch"]!="x86_64", supported_platforms());
       platforms_x86_64 = expand_microarchitectures(platforms_x86_64);
       platforms_x86_64 = filter(p -> p.tags["march"] != "avx512", platforms_x86_64);
       platforms = vcat(platforms_x86_64, platforms_other)

26-element Vector{Platform}:
 Linux x86_64 {libc=glibc, march=x86_64}
 Linux x86_64 {libc=glibc, march=avx}
 Linux x86_64 {libc=glibc, march=avx2}
 Linux x86_64 {libc=musl, march=x86_64}
 Linux x86_64 {libc=musl, march=avx}
 Linux x86_64 {libc=musl, march=avx2}
 macOS x86_64 {march=x86_64}
 macOS x86_64 {march=avx}
 macOS x86_64 {march=avx2}
 FreeBSD x86_64 {march=x86_64}
 FreeBSD x86_64 {march=avx}
 FreeBSD x86_64 {march=avx2}
 Windows x86_64 {march=x86_64}
 Windows x86_64 {march=avx}
 Windows x86_64 {march=avx2}
 Linux i686 {libc=glibc}
 Linux aarch64 {libc=glibc}
 Linux armv6l {call_abi=eabihf, libc=glibc}
 Linux armv7l {call_abi=eabihf, libc=glibc}
 Linux powerpc64le {libc=glibc}
 Linux i686 {libc=musl}
 Linux aarch64 {libc=musl}
 Linux armv6l {call_abi=eabihf, libc=musl}
 Linux armv7l {call_abi=eabihf, libc=musl}
 macOS aarch64
 Windows i686

julia> expand_microarchitectures(supported_platforms(), ["x86_64", "avx", "avx2"])
26-element Vector{Platform}:
 Linux i686 {libc=glibc}
 Linux x86_64 {libc=glibc, march=x86_64}
 Linux x86_64 {libc=glibc, march=avx}
 Linux x86_64 {libc=glibc, march=avx2}
 Linux aarch64 {libc=glibc}
 Linux armv6l {call_abi=eabihf, libc=glibc}
 Linux armv7l {call_abi=eabihf, libc=glibc}
 Linux powerpc64le {libc=glibc}
 Linux i686 {libc=musl}
 Linux x86_64 {libc=musl, march=x86_64}
 Linux x86_64 {libc=musl, march=avx}
 Linux x86_64 {libc=musl, march=avx2}
 Linux aarch64 {libc=musl}
 Linux armv6l {call_abi=eabihf, libc=musl}
 Linux armv7l {call_abi=eabihf, libc=musl}
 macOS x86_64 {march=x86_64}
 macOS x86_64 {march=avx}
 macOS x86_64 {march=avx2}
 macOS aarch64
 FreeBSD x86_64 {march=x86_64}
 FreeBSD x86_64 {march=avx}
 FreeBSD x86_64 {march=avx2}
 Windows i686
 Windows x86_64 {march=x86_64}
 Windows x86_64 {march=avx}
 Windows x86_64 {march=avx2}

julia> all(in(platforms), expand_microarchitectures(supported_platforms(), ["x86_64", "avx", "avx2"]))
true

[ludvigak]

There, I think we are good to go!

[giordano]

Awesome! While we have had for a while this infrastructure to target different microarchitectures, we haven't used it yet into the wild, as you could guess from the aarch64 bugs. This is the first package that is going to actually use it. Should you have any problems, please do let us know!