Web lists-archives.com

Re: RFC: Naming convention for ILP64 variant of BLAS/LAPACK




Hi Wookey and Bastian,

On Sun, Oct 21, 2018 at 06:51:16PM +0100, Wookey wrote:
> On 2018-10-21 17:16 +0200, Bastian Blank wrote:
> > Hi
> > 
> > On Sun, Oct 21, 2018 at 09:51:15AM +0000, Mo Zhou wrote:
> > > about naming convention of SONAME and package name.
> > > 
> > > As discussed in [1][2][3], Debian will need a set of ILP64[4] interface
> > > to BLAS/LAPACK in the future.
> > 
> > Could you please describe what you mean?  All 64-bit Debian
> > architectures are LP64.  So building a single binary using ILP64 will
> > even break the ABI for glibc and it will most likely not run very far.
> > (A file descriptor is defined as "int", so even the most basic file
> > calls will be incompatible.)
 
I missed some points in the original post. The proposal meant to add
a new set of BLAS/LAPACK packages that are compiled in ILP64, and the
existing BLAS/LAPACK API/ABI won't be changed and nothing will break.

Here is a detailed example for demonstration

  src:openblas
    bin:libopenblas-base (LP64, won't be changed at all)
      provides libblas.so.3
	bin:libopenblas-dev (LP64, won't be changed at all)
	  provides libblas.so
	bin:libopenblas-ilp64 (ILP64, newly-added)
	  provides libblas-ilp64.so.3
	bin:libopenblas-ilp64-dev (ILP64, newly-added)
	  provides libblas-ilp64.so

So there is no "ABI bump" but just "adding a new set of ABI". At the
same time, it won't result in any transition or breakage.

> I wondered about this. The mail said that the BLAS/LAPACK ABIs do not
> change, so I presumed that this was about internal data layouts for
> the data being passed which. But reading the bugreps it does sound
> like just a new ABI using ILP64. That would be properly done using
> multiarch or multilib paths, and needs some thought about how best to
> lay things out and what else would be needed to make it work.

The simplest solution is just to create a separate ABI with different
name and different package name. And it introduces the least overhead.

> Are any
> other packages likely to start wanting to use ILP64 ABIs? I guess it's
> very much an 'HPC' sort of thing at the moment.
> 
> So yeah, some clarification in order I think, and an explanation of use-cases.

HPC is indeed a related use case. I don't know any other package that
would need such an ILP64 BLAS/LAPACK interface except for Julia.
Actually by default Julia uses ILP64 version of openblas instead of
LP64, see [julia-ilp64-default].

Here are some references:

1. https://software.intel.com/en-us/mkl-linux-developer-guide-using-the-ilp64-interface-vs-lp64-interface

   The Intel MKL ILP64 libraries use the 64-bit integer type (necessary
   for indexing large arrays, with more than 231-1 elements), whereas
   the LP64 libraries index arrays with the 32-bit integer type.

2.  https://bugzilla.redhat.com/show_bug.cgi?id=1287541

   Julia upstream and fedora chose a different solution... they mangled
   the BLAS/LAPACK symbol names by adding an "64_" suffix. However,
   Julia is the only upstream that use this mangling rule in practice,
   and maybe some other ILP64-capable BLAS/LAPACK providers doesn't
   allow easy name mangling. That means if we still want to take
   advantage from the update-alternatives mechanism, we should not
   mangle the symbol names.

   src:intel-mkl is currently the only ILP64-ready BLAS/LAPACK provider
   in the archive, and it doesn't mangle symbol names.
   

[julia-ilp64-default] https://salsa.debian.org/julia-team/julia/blob/master/DISTRIBUTING.md#L126-130