Skip to content

MPLAPACK 2.3.0

Latest
Latest

Choose a tag to compare

View all tags
@nakatamaho nakatamaho released this 29 Jul 02:34
v2.3.0
4ca5d3d

MPLAPACK 2.3.0 - Changes since 2.2.1

Release date: 2026-07-29 (base commit: afb9637445f980b6ab483f826c9244bac67afb49)

MPLAPACK 2.3.0 is the unified-library release. The public library layout now
builds one self-contained MPLAPACK library per flavor, containing both MPBLAS
and MPLAPACK symbols. This is an intentional ABI and link-interface break from
the 2.2.x layout.

This summary is based on 23 commits in v2.2.1..afb9637445f9.

Unified library layout

  • Every backend/flavor is now self-contained:
    libmplapack_<backend>, libmplapack_<backend>_opt,
    libmplapack_dd_opt_cuda, and
    libmplapack_binary128_opt_opencl.
  • Consumers link exactly one unified flavor library. The separate
    libmpblas_* products and the aggregate mplapack.pc file were removed.
  • Per-flavor pkg-config files and CMake export targets describe the exact
    library and precision dependencies for each flavor.
  • Duplicate public symbols are eliminated by a deterministic basename
    shadowing rule: accelerator sources override optimized sources, and
    optimized sources override reference sources.
  • Archive duplicate-symbol checks are part of the build/test path. The checker
    now handles ELF, Mach-O weak external COMDAT/header-inline symbols, and
    MinGW/PE-COFF archives using a relocatable whole-archive link check.
  • The optimized LAPACK layer is represented explicitly under
    mplapack/optimized/<backend>, even where it is currently empty, so the
    reference, optimized, and accelerator layers stay visible in the source tree.

Version and ABI

  • Package version: 2.3.0.
  • The incompatible library-layout change advances libtool version-info from
    1:0:0 to 2:0:0.
  • CMake shared libraries use version 2.0.0 and SONAME 2.
  • Existing applications that linked libmpblas_* and libmplapack_*
    separately must update link lines to use one unified libmplapack_* flavor.

Autotools and CMake

  • Autotools and CMake now both build the unified reference, optimized, and
    accelerator flavor libraries.
  • CMake provides build-tree and installed targets as
    mplapack::mplapack_<flavor>.
  • MPLAPACK_ENABLE_OPT controls optimized CPU flavors.
  • MPLAPACK_ENABLE_CUDA enables the self-contained DD CUDA flavor.
  • MPLAPACK_ENABLE_OPENCL enables the self-contained binary128 OpenCL flavor.
  • MPLAPACK_CUDA_ARCHITECTURES selects CUDA device architectures.
  • Compare-test generation was updated for unified libraries, and stale
    libmpblas references are rejected.
  • The old compare Makefile generator was removed.
  • The source-manifest drift check compares generated source manifests against
    the reference tree and is registered in both autotools and CTest paths.
  • DD compilation continues to preserve -ffp-contract=off on relevant
    translation units.

Accelerator flavors

  • libmplapack_dd_opt_cuda is now self-contained and links without separate
    MPBLAS/MPLAPACK reference libraries.
  • libmplapack_binary128_opt_opencl is now self-contained and follows the same
    accelerator-over-optimized-over-reference source shadowing rule.
  • The OpenCL Rgemm source was renamed to match the public basename used by
    the duplicate-symbol filter.
  • CUDA builds include a workaround for the glibc C23 rsqrt namespace issue
    seen with newer CUDA toolchains.

Test harness and numerical notes

  • The lin/eig test harnesses are generated for optimized backends in both build
    systems.
  • Standard lin/eig test registration uses the non-optimized input set, so
    release tarballs do not require se2_opt.in for the standard eig harness.
  • The reference-double se2 run remains at threshold 50 and passes.
  • The optimized-double se2 case was reproduced at test 36, matrix type 15,
    N=20, seed 60812,42731,63787,906: ratio 57.2705 with
    OMP_NUM_THREADS=1 and 15.2737 with the default thread count on GCC 15.2.0.
    This is treated as bounded summation-order divergence in the optimized BLAS
    path, so optimized se2 input uses a documented threshold of 100; kernels are
    unchanged.

Release and QA tooling

  • The MPLAPACK test-result summarizer now accepts multiple result roots, prints
    triplet information, and can emit grouped text, CSV, and JSON views.
  • Tier2 and tarball release logs use explicit namespaces matching the tier1
    style.
  • Debian Dockerfiles use noninteractive apt/tzdata setup.
  • openSUSE Leap build dependencies were refreshed, with ppc64le omitted from
    the Leap 15 matrix because binary128 support is unavailable there.
  • macOS and MinGW duplicate-symbol checks were hardened for their native object
    and archive formats.
  • README, README.cmake, migration notes, examples, benchmarks, packaging files,
    and repo-local agent documentation were updated for the unified layout.

2.3.0 test result summary

The 2.3.0 result trees were summarized from:

  • mplapack/test/eig/results/2.3.0/
  • mplapack/test/lin/results/2.3.0/

The result set covers 11 platform/compiler triplets:

  • Apple_M4_macos26_gcc-15_2_0
  • Apple_M4_ubuntu24_04_gcc-13_3_0
  • Apple_M4_ubuntu26_04_gcc-15_2_0
  • Core_i7-6920HQ_macos15_gcc-15_2_0
  • Ryzen_Threadripper_3970X_debian12_gcc-12_2_0
  • Ryzen_Threadripper_3970X_debian13_gcc-14_2_0
  • Ryzen_Threadripper_3970X_ubuntu24_04_gcc-13_3_0
  • Ryzen_Threadripper_3970X_ubuntu24_04_icx-2026_1_0
  • Ryzen_Threadripper_3970X_ubuntu26_04_gcc-15_2_0
  • Ryzen_Threadripper_3970X_ubuntu26_04_icx-2026_0_0
  • Ryzen_Threadripper_3970X_windows_gcc-13_x_x

Aggregate result counts:

Category Triplets Groups .out files Recognized tests Failed tests Failed records
eig 11 96 4,032 116,310,045 8 6
lin 11 96 384 126,515,712 0 0

lin passed across all summarized triplets. eig had 8 failed tests out of
116,310,045 recognized tests, an overall failure rate of about 0.0000069%.

eig failure records:

Triplet Precision File Suite Failed/Total Fail %
Ryzen_Threadripper_3970X_debian12_gcc-12_2_0 double double/Rse2.out DST 1/4,440 0.0225%
Ryzen_Threadripper_3970X_debian13_gcc-14_2_0 double double/Rse2.out DST 1/4,440 0.0225%
Core_i7-6920HQ_macos15_gcc-15_2_0 binary80 binary80/Rsep.out DST drivers 2/13,464 0.0149%
Ryzen_Threadripper_3970X_windows_gcc-13_x_x binary80 binary80/Rsep.out DST drivers 2/13,464 0.0149%
Ryzen_Threadripper_3970X_debian12_gcc-12_2_0 double double/Rsvd.out DBD 1/10,260 0.0097%
Ryzen_Threadripper_3970X_debian13_gcc-14_2_0 double double/Rsvd.out DBD 1/10,260 0.0097%

Audit note: most eig groups had 32 recognized result files. The following
Debian groups had 31 recognized result files in the summarizer output:
Debian 12 double, Debian 12 gmp, Debian 13 dd, Debian 13 double, and
Debian 13 gmp. All lin groups had recognized result files and no failures.

The failure summary can be reproduced from mplapack/test with:

python3 ../../misc/summarize_mplapack_tests.py eig eig/results/2.3.0/* --only-fail
python3 ../../misc/summarize_mplapack_tests.py lin lin/results/2.3.0/* --only-fail
Assets 3
Loading