Index | Thread | Search

From:
Jeremie Courreges-Anglas <jca@wxcvbn.org>
Subject:
Re: numpy 1.24.1 -> 1.25.2
To:
Daniel Dickman <didickman@gmail.com>
Cc:
ports@openbsd.org
Date:
Mon, 08 Jan 2024 16:40:01 +0100

Download raw body.

Thread
  • Jeremie Courreges-Anglas:

    numpy 1.24.1 -> 1.25.2

  • Theo Buehler:

    numpy 1.24.1 -> 1.25.2

  • On Mon, Jan 01 2024, Daniel Dickman <didickman@gmail.com> wrote:
    > The next version of matplotlib requires numpy 1.25 or newer.
    >
    > The below diff updates numpy to 1.25.2 so matplotlib can be updated.
    >
    > Numpy 1.26 switched to a new build system using mesonpy, so someone will 
    > have to port mesonpy before we can get to newer versions.
    >
    > I've tested that all reverse dependencies continue to build on amd64 with 
    > numpy 1.25.2 and run numpy regress tests which look like the usual set of 
    > failures as in older versions:
    >
    > 34 failed, 35312 passed, 1628 skipped, 1308 deselected, 31 xfailed, 4 xpassed, 420 warnings in 405.91s (0:06:45)
    
    -27 failed, 24661 passed, 899 skipped, 1306 deselected, 33 xfailed, 4 xpassed, 49 warnings in 2839.17s (0:47:19)
    +77 failed, 33117 passed, 964 skipped, 1308 deselected, 31 xfailed, 4 xpassed, 410 warnings in 3173.28s (0:52:53)
    
    on riscv64.  It looks worse but all the regressions are in the same
    file/function:
    
      core/tests/test_casting_floatingpoint_errors.py::test_floatingpoint_errors_casting
    
    The issue is most likely that floating-point exceptions are not mandated
    in the RISC-V ISA[0], and have to be implemented by the compiler or in
    the code directly.
    
    [0] https://github.com/riscv-collab/riscv-gnu-toolchain/issues/417
    
    > ok on this update?
    >
    > p.s. Testing on !amd64 is always useful as there has been 
    > platform-specific breakage in the past. Bulk tests would also be helpful 
    > given how much of the tree depends on numpy.
    
    Regarding the bulk build tests: like Stuart I think you do not need more
    than testing that consumers are still happy requirements-wise (which
    I understand you did on amd64).  The positive results of the bulk build
    done by sthen on i386 is quite promising, and so are the additional
    tests on sparc64 and riscv64.  IIUC the only known issue is tb's tests
    results on arm64.  If those are a regression, it would be nice to
    investigate (can't do that myself), if they're not a regression from
    what we have in tree, definitely ok jca@
    
    Logs for riscv64:
    https://wxcvbn.org/~jca/tmp/py3-numpy-1.24.1.log.gz
    https://wxcvbn.org/~jca/tmp/py3-numpy-1.25.2.log.gz
    
    -- 
    jca | PGP : 0x1524E7EE / 5135 92C1 AD36 5293 2BDF  DDCC 0DFA 74AE 1524 E7EE
    
    
    
  • Jeremie Courreges-Anglas:

    numpy 1.24.1 -> 1.25.2

  • Theo Buehler:

    numpy 1.24.1 -> 1.25.2