Minutes_2022_02_15

Numba Meeting: 2022-02-15

Attendees: Siu Kwan Lam, Benjamin Graham, brandon willard, Luk , stuart, Todd A. Anderson (Intel Labs), Val, Ehsan Totoni, Nick Riasanovsky, Guilherme Leobas, Michael Wang,

NOTE: All communication is subject to the Numba Code of Conduct.

Please refer to this calendar for the next meeting date.

0. Feature Discussion

• Numba vision discussion continued

  • Stuart's vision

    • Numba codebase vs project
    • Numba is the default implementation of a compiler toolkit supporting Python and NumPy.
      • covering large spectrum of usecase
      • extensible compiler, scientific usecase, hw vendors
    • Users need to accept that they will need to do some work
    • This kind of flexibility is more important than specific usecase of any group -Luk's question
      • Would Numba the toolkit be split from Numba the NumPy JIT?
      • Stuart's not oppose to this idea

  • Ehsan: next step for defining the vision

    • "A flexible compiler toolkit..."
    • Val: a smaller team to continue defining the vision in separate meeting and report back here
    • Val will drive the vision document with Ehsan and Luk
    • Document principals and values (this isn't the roadmap)

  • Stuart's realistic roadmap

    • important maintaince tasks for new releases of Python, NumPy LLVM
    • ^ reason for more cautious development
    • Now Numba is "HPC Python furniture"
    • upcoming releases:
      • 0.56 June/July - NumPy 1.22, LLVM 13
      • 0.57 Oct/Nov - Py 3.11
      • details in prev notes: https://github.com/numba/numba/wiki/Minutes_2022_01_18
    • ... (more in stuart's note)

  • Nick's suggestions:

    • list underdevelopment areas
    • list high-priority tasks

  • Stuart's Numba the Project

    • people, infrastructure, funding
    • big code base; hard to get started
    • high maintenance burden
    • other needs related to the compiler toolkit
      • profiler
      • coverage tools
      • debugger --- cross hw, cross language (Python interpreted code/JIT compiled Pyython/C)
      • numba-scipy
      • numba-extras
    • values:
      • meritocracy
      • ...

  • Last week's ending questions:

    • What is the core?
    • How to decide what should be in the core?
    • How can we make it easier for people to contribute?

  • Next week we will continue to talk about the decision making progress (e.g. right to merge)

1. New Issues

• #7818 - Calling f2py wrapped code (add an example to documentation)
• #7819 - Reduce inspection of args during kernel launch time
• #7820 - numpy array inconsistency
• #7822 - Fix M1 Unit Test Failures.
• #7824 - _helperlib.c::import_cython_function erroneously checks result of PyCapsule_GetName against NULL
• #7830 - Installing numpy and numba with conda-forge results in apparently incompatible versions on Python 3.8 and 3.9
• #7831 - Investigation in Dispatcher immortality
• #7832 - manage_memory() needs a param for extra dtor data
• #7834 - Support passing non-constant values to Exceptions in JIT
• #7838 - Array analysis pass is not idempotent
• #7839 - Using named tuples of type typing.NamedTuple crashes the kernel
• #7840 - Order-preserving slice of a Fortran-ordered array results in an A-ordered array
• #7841 - Compilation error under parfor in parallel decorator
• #7842 - jit/njit parallelization of np.roll()
• #7843 - Incorrect result when numpy broadcasting in parallel

Closed Issues

• #7821 - segfault after setting numba.config.NUMBA_NUM_THREADS

2. New PRs

• #7823 - Add renamed vars to callee scope such that it is self consistent.
• #7825 - F2PY Support
• #7826 - Feature/add support for numpy insert
• #7827 - Overload array
• #7828 - Set maximum name size to maximum allowable value
• #7829 - CUDA: Support Enum/IntEnum in Kernel
• #7833 - Add version support information table to docs.
• #7835 - [WIP] Fix pickling error when module cannot be imported
• #7836 - Datetime min max
• #7837 - Initial refactoring of parfor reduction lowering

Closed PRs

3. Next Release: Version 0.56.0/0.39.0, RC,