On Tue, 2019-05-28 at 01:37 -0700, Mo Zhou wrote:
> Hi Gentoo devs,
>
> Classical numerical linear algebra libraries, BLAS[1] and LAPACK[2]
> play important roles in the scientific computing field, as many
> software such as Numpy, Scipy, Julia, Octave, R are built upon them.
>
> There is a standard implementation of BLAS and LAPACK, named netlib
> or simply "reference implementation". This implementation had been
> provided by gentoo's main repo. However, it has a major problem:
> performance. On the other hand, a number of well-optimized
> BLAS/LAPACK
> implementations exist, including OpenBLAS (free), BLIS (free),
> MKL (non-free), etc., but none of them has been properly integrated
> into the Gentoo distribution.
>
> I'm writing to propose a good solution to this problem. If no gentoo
> developer is object to this proposal, I'll keep moving forward and
> start submitting PRs to Gentoo main repo.
>
> Historical Obstacle
> -------------------
>
> Different BLAS/LAPACK implementations are expected to be compatible
> to each other in both the API and ABI level. They can be used as
> drop-in replacement to the others. This sounds nice, but the
> difference
> in SONAME hampered the gentoo integration of well-optimized ones.
>
> Assume a Gentoo user compiled a pile of packages on top of the
> reference
> BLAS and LAPACK, namely these reverse dependencies are linked against
> libblas.so.3 and liblapack.so.3 . When the user discovered that
> OpenBLAS provides much better performance, they'll have to recompile
> the whole reverse dependency tree in order to take advantage from
> OpenBLAS,
> because the SONAME of OpenBLAS is libopenblas.so.0 . When the user
> wants to try MKL (libmkl_rt.so), they'll have to recompile the whole
> reverse dependency tree again.
>
> This is not friendly to our earth.
>
> Goal
> ----
>
> * When a program is linked against libblas.so or liblapack.so
> provided by any BLAS/LAPACK provider, the eselect-based solution
> will allow user to switch the underlying library without
> recompiling
> anything.
>
> * When a program is linked against a specific implementation, e.g.
> libmkl_rt.so, the solution doesn't break anything.
>
> Solution
> --------
>
> Similar to Debian's update-alternatives mechanism, Gentoo's eselect
> is good at dealing with drop-in replacements as well. My preliminary
> investigation suggests that eselect is enough for enabling
> BLAS/LAPACK
> runtime switching. Hence, the proposed solution is eselect-based:
>
> * Every BLAS/LAPACK implementation should provide generic library
> and eselect candidate libraries at the same time. Taking netlib,
> BLIS and OpenBLAS as examples:
>
> reference:
>
> usr/lib64/blas/reference/libblas.so.3 (SONAME=libblas.so.3)
> -- default BLAS provider
> -- candidate of the eselect "blas" unit
> -- will be symlinked to usr/lib64/libblas.so.3 by eselect
>
> usr/lib64/lapack/reference/liblapack.so.3
> (SONAME=liblapack.so.3)
> -- default LAPACK provider
> -- candidate of the eselect "lapack" unit
> -- will be symlinked to usr/lib64/liblapack.so.3 by eselect
>
> blis (doesn't provide LAPACK):
>
> usr/lib64/libblis.so.2 (SONAME=libblis.so.2)
> -- general purpose
>
> usr/lib64/blas/blis/libblas.so.3 (SONAME=libblas.so.3)
> -- candidate of the eselect "blas" unit
> -- will be symlinked to usr/lib64/libblas.so.3 by eselect
> -- compiled from the same set of object files as libblis.so.2
>
> openblas:
>
> usr/lib64/libopenblas.so.0 (SONAME=libopenblas.so.0)
> -- general purpose
>
> usr/lib64/blas/openblas/libblas.so.3 (SONAME=libblas.so.3)
> -- candidate of the eselect "blas" unit
> -- will be symlinked to usr/lib64/libblas.so.3 by eselect
> -- compiled from the same set of object files as
> libopenblas.so.0
>
> usr/lib64/lapack/openblas/liblapack.so.3
> (SONAME=liblapack.so.3)
> -- candidate of the eselect "lapack" unit
> -- will be symlinked to usr/lib64/liblapack.so.3 by eselect
> -- compiled from the same set of object files as
> libopenblas.so.0
>
> This solution is similar to Debian's[3]. This solution achieves our
> goal,
> and it requires us to patch upstream build systems (same to Debian).
> Preliminary demonstration for this solution is available, see below.
>
> Is this solution reliable?
> --------------------------
>
> * A similar solution has been used by Debian for many years.
> * Many projects call BLAS/LAPACK libraries through FFI, including
> Julia.
> (See Julia's standard library: LinearAlgebra)
>
> Proposed Changes
> ----------------
>
> 1. Deprecate sci-libs/{blas,cblas,lapack,lapacke}-reference from
> gentoo
> main repo. They use exactly the same source tarball. It's not
> quite
> helpful to package these components in a fine-grained manner. A
> single
> sci-libs/lapack package is enough.
>
> 2. Merge the "cblas" eselect unit into "blas" unit. It is potentially
> harmful when "blas" and "cblas" point to different
> implementations.
> That means "app-eselect/eselect-cblas" should be deprecated.
>
> 3. Update virtual/{blas,cblas,lapack,lapacke}. BLAS/LAPACK providers
> will be registered in their dependency information.
>
> Note, ebuilds for BLAS/LAPACK reverse dependencies are expected to
> work
> with these changes correctly without change. For example, my local
> numpy-1.16.1 compilation was successful without change.
>
> Preliminary Demonstration
> -------------------------
>
> The preliminary implementation is available in my personal
> overlay[4].
> A simple sanity test script `check-cpp.sh` is provided to illustrate
> the effectiveness of the proposed solution.
>
> The script `check-cpp.sh` compiles two C++ programs -- one calls
> general
> matrix-matrix multiplication from BLAS, while another one calls
> general
> singular value decomposition from LAPACK. Once compiled, this script
> will switch different BLAS/LAPACK implementations and run the C++
> programs
> without recompilation.
>
> The preliminary result is avaiable here[5]. (CPU=Power9,
> ARCH=ppc64le)
> From the experimental results, we find that
>
> For (512x512) single precision matrix multiplication:
> * reference BLAS takes ~360 ms
> * BLIS takes ~70 ms
> * OpenBLAS takes ~10 ms
>
> For (512x512) single precision singular value decomposition:
> * reference LAPACK takes ~1900 ms
> * BLIS (+reference LAPACK) takes ~1500 ms
> * OpenBLAS takes ~1100 ms
>
> The difference in computation speed illustrates the effectiveness of
> the proposed solution. Theoretically, any other package could take
> advantage from this solution without any recompilation as long as
> it's linked against a library with SONAME.
>
> Acknowledgement
> ---------------
> This is an on-going GSoC-2019 Porject:
> https://summerofcode.withgoogle.com/projects/?sp-page=2#6268942782300160
>
> Mentor: Benda Xu
>
> [1] BLAS = Basic Linear Algebra Subroutines. It's a set of API + ABI.
> [2] LAPACK = Linear Algebra PACKage. It's a set of API + ABI.
> [3] https://wiki.debian.org/DebianScience/LinearAlgebraLibraries
> [4] https://github.com/cdluminate/my-overlay
> [5]
> https://gist.github.com/cdluminate/0cfeab19b89a8b5ac4ea2c5f942d8f64
>
We already have such a solution in the sci-overlay. It has proven
extremely brittle and shaky. The plan is to do this via USE flags
similar to python-single-r1 flags. Optionally, we can leave a "virtual"
USE flag, where users can switch implementation at runtime, but this
will not be a supported configuration.
While I understand that runtime-switching sounds like a great feature,
it has proven too difficult to get right and too hard to enforce
invariants on correct symlinks. People that want this can go the
virtual+eselect approach in the overlay, but 99% of Gentoo users will
be happy with just linking against OpenBLAS/reference-lapack and not
having to fix weird stale symlinks that eselect-alternatives somehow
lost track of.
David
See also:
https://bugs.gentoo.org/632624
https://bugs.gentoo.org/348843#c30
