Dear Fabian,
Thank you for your answer. I assume there is no way to leave the AVX2
support restricted to the amd64 build only then.
I do really appreciate all the efforts to get programs like shapeit4 in a
debian package format. It is a great gift to the community of
bioinformaticians.
I'll see if I can get the author to take on your suggestion of including
the selection of AVX2 at runtime, which I think is the right approach.
However, in my previous experience the author has been quite non-responsive.
Giulio
On Thu, Nov 5, 2020 at 2:46 PM Fabian Klötzl <kloe...@evolbio.mpg.de> wrote:
> Hi all,
>
> I agree that most CPUs used for scientific computing these days support
> AVX2. However, that is going to change. For instance, Apple is switching
> to ARM based CPUs. MacBooks are quite popular among bioinformaticians so
> so any tool author would want to support non-x86 CPUs. I also believe
> that other manufacturers will follow Apples lead to Arm.
>
> That said, how can one support different instruction sets and get the
> optimal performance? Well, depending on the code, it can be simple or hard.
>
> If there is just one function which needs the boost from the instruction
> set, one can simply add an __attribute__((target_clones("…"))) [1]
>
> However, one can also compile different sections of the code with
> different optimizations and then use __builtin_cpu_supports() to decide
> at runtime which version should be used. I used that strategy with
> ifuncs for phylonium [2]. This way is more difficult and I haven't seen
> a good writeup of it yet.
>
> Unfortunately both ways require quite a bit work from the upstream
> author. If someone knows a better way, let me know.
>
> Best,
> Fabian
>
>
> 1: https://lwn.net/Articles/691932/
> 2: https://salsa.debian.org/med-team/phylonium
>
>
>
> On 05.11.20 15:02, Dylan Aïssi wrote:
> > Dear Giulio,
> >
> > I am CCing the public Debian Med mailing list and Michael Crusoe who
> > can help to improve the Debian package in order to have an AVX2 binary
> > optimized.
> >
> > Le jeu. 5 nov. 2020 à 15:20, Giulio Genovese
> > <giulio.genov...@gmail.com> a écrit :
> >>
> >> I have noticed you are the author of the following patch for the
> shapeit4 debian package:
> >>
> https://salsa.debian.org/med-team/shapeit4/-/blob/master/debian/patches/use_shared_libs.patch
> >>
> >> I noticed that this causes shapeit4 to not use the AVX2 instruction set:
> >> -CXXFLAG=-O3 -mavx2 -mfma
> >> +#CXXFLAG=-O3 -mavx2 -mfma
> >> #Portable version without avx2 (much slower)
> >> -#CXXFLAG=-O3
> >> -LDFLAG=-O3
> >> +CXXFLAG=$(CPPFLAGS) $(CXXFLAGS) -O3
> >> +LDFLAG=$(LDFLAGS) -O3
> >>
> >> This causes shapeit4 to run significantly slower than it could on most
> systems where this package would be installed.
> >>
> >> I assume that the reason for this is that older CPUs do not support
> AVX2. Is it really important though to support these old systems? I doubt
> anybody would run such a tool on old machines. However, there might be
> reasons for this decision that I might be missing so I am curious to know
> the goals of this modification.
> >>
> >
> > You are right, this is the Debian policy to provide binaries that
> > respect the architecture baseline [1] to support older CPUs. And you
> > are also right when you said nobody uses this tool on this kind of
> > outdated CPU. Personally, I don't have time to work on that but I
> > guess someone else in the Debian Med team would be interested?
> >
> > Best,
> > Dylan
> >
> > [1] https://wiki.debian.org/ArchitectureSpecificsMemo#amd64
> >
>
