[bug#52283] [PATCH 00/10] Tuning packages for CPU micro-architectures

[Ludovic Courtès]

Hello Guix!

This patch series is an attempt to allow users to build or
substitute packages for the very CPU they are using, as opposed
to using a generic binary that targets the baseline
architecture—e.g., x86_64 without AVX extensions.

As a reminder, my take on this is that The Right Thing is for
code to select optimized implementations for the host CPU at
load time, using (possibly hand-crafted) “function multi-versioning”:

  https://hpc.guix.info/blog/2018/01/pre-built-binaries-vs-performance/

Now, there’s at least one situation where developers don’t do
“the right thing”: C++ header-only libraries.  It turns out
header-only libraries with #ifdef’d SIMD code are quite common:
Eigen, xsimd, xtensor, etc.  Every user of those libs has to be
compiled with ‘-march=native’ to take advantage of those
SIMD-optimized routines and there’s little hope of seeing those
libraries implement load-time or run-time selection¹.

This patch set implements “package multi-versioning”, where a package
can have different variants users may choose from: baseline, haswell,
skylake, etc.  This is implemented as a package transformation option,
‘--tune’.  Without any argument, ‘--tune’ grafts tuned package variants
for each package that has the ‘tunable?’ property.  For example:

  guix shell eigen-benchmarks --tune -- benchBlasGemm 16 16 16 100 100

runs one of the Eigen benchmarks tuned for the host CPU, because
‘eigen-benchmarks’ is marked as “tunable”.

This is achieved not by passing ‘-march=native’, because the daemon
might be running on a separate machine with a different CPU, but by
identifying the ‘-march’ value corresponding to the host CPU and
passing ‘-march’ to the compiler, via a wrapper.

On my skylake laptop, that gives a noticeable difference on the GEMM
benchmark of Eigen and good results on the xtensor benchmarks too,
unsurprisingly.  I don’t have figures for higher-level applications,
but it’d be nice to benchmark some of Eigen’s dependents for instance,
as shown by:

  guix graph -M2 -t reverse-package eigen | xdot -f fdp -

If you could run such benchmarks, that’d be great!  :-)
Things like Fenics may benefit from it.

Nix people chose to introduce separate system types for the various
x86_64 micro-architecture levels: x86_64-linux-v1, x86_64-linux-v2,
etc.²  I think this is somewhat wasteful and unpractical though.
It’s also unclear whether those levels, defined in the new x86_64
psABI³, are a viable abstraction: vendors seem to be mixing features
rather than really following the accumulative pattern that those
levels imply.

Thoughts?

Ludo’.

¹ 
https://listengine.tuxfamily.org/lists.tuxfamily.org/eigen/2021/11/msg00006.html
² https://discourse.nixos.org/t/nix-2-4-released/15822
³ 
https://gitlab.com/x86-psABIs/x86-64-ABI/-/blob/master/x86-64-ABI/low-level-sys-info.tex

Ludovic Courtès (10):
  Add (guix cpu).
  transformations: Add '--tune'.
  ci: Add extra jobs for tunable packages.
  gnu: Add eigen-benchmarks.
  gnu: Add xsimd-benchmark.
  gnu: Add xtensor-benchmark.
  gnu: ceres-solver: Mark as tunable.
  gnu: Add ceres-solver-benchmarks.
  gnu: libfive: Mark as tunable.
  gnu: prusa-slicer: Mark as tunable.

 Makefile.am                  |   1 +
 doc/guix.texi                |  54 ++++++++++++++
 gnu/ci.scm                   |  43 ++++++++---
 gnu/packages/algebra.scm     |  79 ++++++++++++++++++++
 gnu/packages/cpp.scm         |  23 ++++++
 gnu/packages/engineering.scm |  10 ++-
 gnu/packages/maths.scm       |  49 ++++++++++++-
 guix/cpu.scm                 | 137 +++++++++++++++++++++++++++++++++++
 guix/transformations.scm     | 134 ++++++++++++++++++++++++++++++++++
 tests/transformations.scm    |  20 +++++
 10 files changed, 538 insertions(+), 12 deletions(-)
 create mode 100644 guix/cpu.scm


base-commit: 052f56e5a614854636563278ee5a2248b3609d87
prerequisite-patch-id: 7e5c2bb5942496daf01a7f6dfc1b0b5b214f1584
-- 
2.33.0