[Debian-med-packaging] Help to build library in generic form, avx and sse3

[Christian Seiler]

On 03/14/2017 03:46 PM, Andreas Tille wrote:
> I've started packaging Phylogenetic Likelihood Library[1].  Since it
> makes heavy use of amd64 features it comes with specific support of AVX
> and SSE3.  My plan is to provide binary packages amd64 only named
> libpll-avx1 and libpll-sse3-1 with the according features plus a generic
> library libpll-generic1 for all architectures.  Upstream supports the
> creation of separate avx and sse3 libs out of the box but I failed to
> create the generic version.  So I have two questions:
> 
>   1. Could anybody please have a look at the automake stuff to
>      enable the build of the generic lib in addition to the other
>      two.  I tried several switches but failed. :-(
> 
>   2. What do you think about the plan to support specific hardware
>      features in separate binary packages?

GCC from version 6 (which is in Debian Stretch) supports function
multi-versioning (and GCC from 4.8 onwards, which is even in Jessie,
supports a subset of that), which allows you to do the following:

 - have a function with generic C/C++ code be compiled multiple
   times in different variants, and have the most optimal variant
   be selected at runtime (requires GCC 6),

   e.g.

   __attribute__((target_clones("avx2","sse3","default")))
   double foo(double a, double b) {
     return a + b;
   }

 - manually write different versions of the function and mark
   them accordingly (requires GCC 4.8)

   __attribute__((target("default")))
   double foo(double a, double b) {
     return a + b;
   }

   __attribute__((target("sse3")))
   double foo(double a, double b) {
     SOME_FANCY_SSE3_CODE;
   }

   __attribute__((target("avx2")))
   double foo(double a, double b) {
     SOME_FANCY_AVX2_CODE;
   }

So from a purely technical perspective I think the best solution
would probably be to work with upstream to allow them to support
FMV properly - and then you only need to compile a single library
version that will work everywhere, but will select the optimal
algorithm depending on the machine it's run - win/win.

Further reading:
https://lwn.net/Articles/691932/

Regards,
Christian