Skip to content

MPLAPACK 2.2.1

Choose a tag to compare

View all tags
@nakatamaho nakatamaho released this 21 Jul 03:11
· 123 commits to master since this release
v2.2.1
8fb8c83

MPLAPACK 2.2.1 - Changes since 2.2.0

Release date: 2026-07 (base commit: 8fb8c83427c3561e003fab19202cd8c0f7833bbb)

MPLAPACK 2.2.1 is a patch release on top of 2.2.0. It keeps the
LAPACK 3.12.1-based reference code from 2.2.0 and focuses on GMP
transcendental correctness, complex helper coverage, binary128/binary80
C++ standard compatibility, OpenCL binary128 work, an experimental CMake
build, and release engineering for the expanded build matrix.
This summary is based on 168 commits in v2.2.0..8fb8c834.

No public BLAS/LAPACK routine set is intentionally changed in this
release. The notable user-visible changes are in backend helper behavior,
headers, configure/CMake feature detection, examples, benchmarks, and
release/test infrastructure.

1. GMP backend transcendental functions

  • Added native GMP-backed helpers for log2, log1p, log, exp,
    expm1, sin/cos, atan, atan2, and pow, with MPFR-free
    primary tests and an MPFR oracle test where appropriate
    (1764beada through 3265bda91, 566680bd4, b1c515a88).
  • Added AGM-based constants and logarithms, plus a quo_rem argument
    reduction primitive used by the trigonometric path.
  • Hardened the transcendental implementation after merge:
    denominator-overflow handling in sincos, tighter quo_rem precision
    handling, expm1 cancellation fixes, and additional edge-case tests
    (4e0325007, f346408eb, 52e3f3aa7, d15693a9a,
    636d8f959).
  • Re-enabled the GMP transcendental path in LATMS/LATMT, enabled GMP
    CSD tests, and removed the temporary GMP-specific BBCSD tolerance
    override (ef3cd6993, 56285b344, 4e9653233).
  • Fixed GMP singleton interval checks in Rstebz (567344252,
    6c8c842f3).
  • Relaxed one GMP cosine-vs-MPFR comparison tolerance on i386
    (41129b5bf).

2. Multiprecision complex and predicate hardening

  • Completed multiprecision complex helper coverage and preserved
    mpreal precision when copying (4c58126d3, 2d963aa24).
  • Added MPC pointer conversions for mpcomplex (a823c3a19).
  • Applied safer mpcomplex constructor behavior (f038f019e).
  • Fixed NaN and infinity predicates across generated headers
    (65ac082fb).
  • Added mpcomplex overloads for binary complex addition and arithmetic
    with std::complex<_Float128> / std::complex<__float128> /
    std::complex<_Float64x> forms, avoiding ambiguous overloads in
    C++23-era builds (4e7992b7c, b54ef2506).

3. C++ standard and binary128/binary80 feature detection

  • Added --with-cxx-standard support for C++17, C++20, C++23, and
    C++26, including GNU-extension and strict-standard modes where
    supported (4957620ca, 3d25873f7).
  • Hardened binary128 _Float128 math selection:
    std::abs, scalar math (sin, cos, atan2, exp, log, etc.),
    std::complex, and C complex support are now controlled by configure
    probes rather than only by language-mode assumptions
    (8f8cdf62e, f5de81ace, b0351c2d5).
  • Added analogous binary80 _Float64x math and abs probing
    (b0351c2d5).
  • Extended configure summaries to report binary128/binary80 std::abs,
    std::math, std::complex, and C-complex availability, including the
    probed function groups.
  • Added Ubuntu 26.04 C++ standard build scripts and Tier 3 C++ standard
    coverage for both amd64 and arm64 (50a181857, e8d16eb99,
    dadfbb45a).
  • Documented the binary128/binary80 type-support and math-support
    situation in binary128_binary80_type_support.md (b0351c2d5,
    2353d0352).

4. Header include-order fixes

  • mplapack_binary128.h now includes <complex> itself so users can
    include the installed LAPACK declarations directly (0786e001b).
  • mpblas_binary128.h was hardened so mplapack_config.h appears before
    standard headers that may depend on feature-test macros
    (2b92d5a21).
  • mpblas_binary80.h received the same include-order hardening
    (308abcccd).

5. Experimental CMake build

  • Added an experimental modern CMake build system with FetchContent-based
    dependency support (073560fa8).
  • Added install/package support, pkg-config/CMake package files, and CPack
    source/binary packaging as the CMake-side equivalent of make dist
    (6d88fee92).
  • Ported the MPBLAS correctness tests and mplapack/test/compare suite to
    CMake, including link-group/whole-archive fixes for the test override
    libraries (19a041488, 7a69e0279, 5f7ecb460).
  • Added CMake benchmark coverage for reference BLAS vs. OpenBLAS
    comparison builds (38a302ddf, d0ded909d).
  • Added CMake-side binary128/binary80 math probes so the CMake config
    header mirrors the autotools probe results for the affected math and
    abs helpers (2353d0352).
  • Documented the experimental CMake build in README.md; the autotools
    build remains the primary release path (2353d0352).

6. OpenCL binary128 work

  • Merged and ported OpenCL support for binary128 RGEMM, including older
    OpenCL platform/device fixes and moving scalar parameters to the GPU
    path (353b090b2, b88af945e, f1a7462be, dc079c806).
  • Removed obsolete _Float128 OpenCL sources after the port
    (2770aec9e).
  • Added binary128 OpenCL benchmark targets and an Rgemm OpenCL benchmark
    runner (c46c82e18, caed88118).
  • Made OpenCL enablement depend on a usable OpenCL configuration and
    gated OpenCL binary128 in remote builds (58e369c27, ba7fae1cf).
  • Used a portable standard-library include in the binary128 OpenCL path
    (70ac89c78).

7. External libraries and toolchain fixes

  • Updated MPC to 1.4.1 and added checksum verification for external
    distfiles (c3daa1f8d, bd4f4e25f).
  • Updated OpenBLAS to 0.3.33, then pruned obsolete external archives and
    retained the required OpenBLAS version in the external tree
    (e64ecc16e, 540fc0f9f, 8ab77298f, 8201aaabc).
  • Fixed Intel oneAPI OpenBLAS configure handling and switched oneAPI
    benchmark BLAS selection to MKL where appropriate (ce438a2a6,
    0ce19284c).
  • Propagated quadmath flags to oneAPI examples and distributed Intel
    oneAPI example makefiles (5cc3a32d3, 1d1616c47, dd31b81ff).
  • Fixed macOS SDK Mach-O include handling with GNU C++ (9d65b086d).
  • Avoided a glibc rsqrt conflict in CUDA builds and generated CUDA
    libtool objects without depending on system libtool
    (59b1637c4, 57593f553).
  • Enforced a single -std=gnu++17 in the legacy configure path so stale
    environment flags do not accumulate conflicting -std= options
    (ef772cc19).

8. Release engineering

  • Version bumped to 2.2.1 in configure.ac (0c4c3c609,
    ee73cd2c6).
  • Release buildtests were moved to remote tier targets, with remote macOS
    and Ubuntu arm64 Docker coverage added (1bf24c411, ee09860db,
    67f90f0c4, dd7f2ba7a).
  • Release target names now include distro versions and use ref
    terminology consistently (cba67fb1d, 2c5f98430).
  • Release jobs run by host and use stable host target files so SSH does
    not consume target lists (d0f833661, 65c8b01c3, 4f4a2ed53).
  • Remote jobs are serialized per host, so make -j tier1 tier2 tier3 tarball can prepare independent targets while only one job runs on a
    given host (fe8f283be).
  • Release source snapshots are now preserved across tier runs, dist
    tarballs can be cached by source key, and fresh source snapshot
    generation was fixed (af13a20b8, e8fc3c424, 05ed567a1).
  • make realclean and remote cleanup were tightened to remove local and
    remote release logs, temporary workdirs, success/failure stamps, and
    QA-tagged Docker containers/images without deleting the local checkout
    (d4fb6eaa7, 8fb8c8342).
  • Release logging now records timings, failures, and ccache statistics
    (a9fd658a5, 0026d55d9, a4a94f266).
  • Release Docker matrix layout was refactored; tarball smoke tests now
    run on remote native Docker hosts and use xz artifacts
    (3ce0deb13, 42350f31f, 0649360b6).
  • Installed examples and benchmark scripts are checked in Tier 1
    release testing, with example targets filtered by enabled backends
    (6c2de4952, 0915833f0).
  • ccache use was made more consistent in release scripts, including
    Fortran reconfig scripts (afc842962, cc3ed8f9c).

9. Build matrix and developer containers

  • Added an Ubuntu 26.04 development container and subsequent tool updates,
    including CUDA/WineHQ setup, clang, and portable clang-format wrappers
    (a5c804d00, cae44a64c, 015612c82, 9eb78d462,
    8820b5441).
  • Refined Tier 1 / Tier 2 / Tier 3 release coverage: Ubuntu amd64 NVIDIA
    distcheck, Ubuntu arm64 Docker, Debian i386 split image, MinGW updates,
    Intel oneAPI entries, and selected native benchmarks on non-x86
    (bca934bf5, 5f6a670d9, 839dae437, 3ab6e8e26).
  • Added branch-matrix and Dockerfile cleanups, including removal of old
    tarball Dockerfiles and stale release distcheck logs (1b78b86e0,
    9084dede2, 5e38066cb).

10. Test results and documentation

  • Added and reorganized 2.2.0 test result logs for
    Core_i5-9600_ubuntu22_04_gcc-11_4_0 and updated README release news
    after the 2.2.0 release (e2cc45c13, a7bf9090d).
  • Added Git checkout bootstrap requirements documentation (f8decdeea).
  • Updated MPLAPACK 3.0 TODO notes (c9957c07d).
  • Ignored new autotools, CMake, and release distcheck artifacts in
    .gitignore (58a65ce5c, b12ae7ecc).

11. Compatibility notes

  • Re-run configure or CMake when changing the C++ language standard.
    The binary128/binary80 math helper decisions are probe results and can
    differ between c++17, gnu++17, c++20, c++23, and c++26.
  • Code that directly includes installed binary128/binary80 BLAS/LAPACK
    headers no longer needs to include <complex> first.
  • The CMake build is still experimental in 2.2.1. Use autotools for the
    primary release build and use the CMake path for targeted validation and
    developer testing.
Assets 5
Loading