Minutes_2022_01_11

Numba Meeting: 2021-1-10

Attendees: Val, Graham Markall, ben@bodo.ai, brandom willard, Ehsan Totoni,, Guilherme Leobas, Luk, Nick Riasanovsky, Todd A. Anderson, stuart

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

Please refer to this calendar for the next meeting date.

0. Feature Discussion

• llvmlite RC/final status

  • going straight to final as no major blockers ahve been been reported.
  • ETA Monday 17th January 2022

• numba RC/final status

  • 4 issues remaining on RC2 milestone
  • ETA Thursday 20th January 2021

• #7717 - discuss next week again when Siu is around

• #7681 Further discussion on Weds 12th Jan

  • Array Analysis requires further functionality / development

• #7701: Feedback on the idea of warning or erroring when nan-unsafe operations done in fastmath mode requested.

  • Similar for isinf?
  • Post on discourse + link to issues + forward issues to discourse

• Compiler explorer for Numba

1. New Issues

• #7653 - Better, more official way to get specific macOS SDKs?
• #7654 - Record mangling_args should not use self._code
• #7655 - parallel=True does not work for functions with a jitclass instance as an argument
• #7657 - Numba 0.55.0rc1 Checklist
• #7658 - parallel LoweringError when aliasing a = a.flat
• #7662 - IntEnum as array index? No implementation of getitem found for signature
• #7663 - Feature request numpy.unique axis=0
• #7665 - Poor np.full performace with jit
• #7666 - numpy.copy messes up result of njitted generator
• #7667 - @jit method return two-dimentional array with wrong shape
• #7668 - aot compile raise error with yield
• #7669 - Immutable error: "No implementation of function Function(<built-in function setitem>) found for signature"
• #7670 - [FEATURE REQUEST] Raise custom exception when users try to import numba.roc
• #7674 - FR: Support numpy.lib.stride_tricks.sliding_window_view
• #7675 - Fail to compile prange with slice in variable resulting in attempt to getitem with wrong number of indices
• #7676 - Incorrect results with prange and sliced array
• #7677 - parallel=True causes AssertionError in slicing even with no parallel transformation
• #7678 - Reflected lists aren't usable as jitclass members
• #7679 - Numba CUDA Matrix Multiplication Example
• **** #7681 - Bug of mismatch shape with parallel mode
• #7683 - connot create array types inside jitted functions
• #7686 - numba and parallel random generators on CPU
• #7693 - Incorrect result copying array-typed field of structured array in 0.55.0rc1
• #7694 - GLIBC version error in 0.55RC1 on AArch64
• #7696 - support for nogil and parallel in AOT functions
• #7697 - CUDA: support recursion in device functions
• #7698 - AssertionError if you don't pass a Signature to cuda.declare_device
• **** #7701 - np.nanmean doesn't ignore nans in fastmath mode
• #7706 - Numba CUDA Array interface faster than Pytorch CUDA Array
• #7707 - Cannot box Literals that aren't hashable
• #7710 - CUDA: Don't use typemap for function name and line number
• #7711 - in operator for the type returned by numba.typed.Dict().keys()?
• #7713 - Tests using float hashing fail on win32+Py3.10 for 0.55rc1
• **** #7716 - Numba doesn't keep a reference to dynamic global array

Closed Issues

• #7652 - "Overwrite of parallel loop index" with inlined functions and numba.prange
• #7656 - Numba fails to install on Windows due to llvmlite error
• #7671 - implementing n-tree is not possible
• #7672 - jitclass with ListType doesn't work
• #7673 - Numpy matrix inside a method
• #7692 - Error installing numba 0.51.2
• #7715 - methods of numpy.random can not reproduce the same probability with the same random seed under the nopython=True

2. New PRs

• #7660 - Add support for np.broadcast_arrays
• #7664 - Flatten mangling dicts into a single dict
• #7680 - CUDA Docs: include example calling slow matmul
• #7682 - performance improvements to np.full and np.ones
• #7685 - Don't convert setitems that have dimension mismatches to parfors.
• #7687 - - remove setuptools dependency
• #7690 - Implemented np.random.noncentral_chisquare for all size arguments
• #7699 - CUDA: Provide helpful error if the return type is missing for declare_device
• #7700 - Support for scalar arguments in Np.ascontiguousarray
• #7703 - Ignore unsupported types in ShapeEquivSet._getnames()
• #7705 - Use omp_set_max_active_levels() instead of omp_set_nested()
• #7708 - Enable Boxing Unhashable Literals
• #7712 - Experiment for issue 7693
• #7714 - Support for boxing SliceLiteral type
• #7717 - [WIP] Keep references to dynamic global arrays

Closed PRs

• #7648 - [TESTING] Try to get error compileing gufunc_scheduler.cpp on OS X
• #7649 - [DO NOT MERGE] Hack/pr/7639
• #7650 - Move Azure builds to OSX 10.15
• #7651 - DOC: pypi and conda-forge badges
• #7659 - Add support for np.broadcast_arrays
• #7661 - Pin packages for release 0.55 branch
• #7684 - DOC: remove incorrect warning in np.random reference
• #7688 - Fix/4927 slow import
• #7689 - Fix/4927 slow import take2
• #7691 - Fix NUMBA_DUMP_ANNOTATION with NUMBA_EXTEND_VARIABLE_LIFETIMES
• #7695 - IntEnumMember support for np.empty, np.zeros, and np.ones
• #7702 - CI testing: Pin NumPy to 1.21 for Python 3.10.
• #7704 - Move the type annotation pass to post legalization.
• #7709 - CUDA: Fixes missing type annotation pass following #7704

3. Next Release: Version 0.55.0/0.38.0, Final