Minutes_2021_03_02

Numba Meeting: 2021-03-02

Attendees: Val, Siu, Ehsan, Graham, Jim, Luk, P-Ortmann, Stuart, Todd

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

0. Feature Discussion/admin

• RC2 status
  • https://github.com/numba/numba/milestone/54
    • one regression so far related gufunc
    • https://github.com/numba/numba/issues/6773
  • numba intergration test
    • status: most likely okay; failure seems unrelated to numba
• Discussion on Roles
  • presentation: https://hackmd.io/fApKIH4QSc22wtTbppZ8Fg
  • Val: how to incentivize maintainers to stay?
  • Ehsan: thinks incentives are there. Problem is to get started to be a maintainer.
  • Val on how to get started:
    • trying fixing issues, become a triager to help other to fix issue.
    • training materials, courses?
  • Ehsan on training:
    • Good high-level arch doc; the flow
      • expand the numba dev docs
    • Good coding convention
      • e.g. flake8, docstrings, ...
      • congnitive complexity; smaller modules/functions
      • less magic; more structures
  • Graham:
    • not lost due to size of things but cross-modules; hard to follow
    • because of the complexity; only high-stake contributors will stay around
  • Siu: restructuring and more interfaces will be needed but the refactoring will take time
  • Stuart: refactoring is hard.
  • Graham: experience on being a Triager is a good way to see different places of numba
  • Stuart: triage pairing/mentor

1. New Issues

• #6774 - parallel bugs with bool arrays?
• #6773 - Regression in 0.53rc2: axis= parameter in gufuncs
• #6772 - Numba 0.53.0rc2 function-as-argument caching fails some of the time
  • function argument caching were never supported.
• #6768 - numpy.arange and numpy.linspace not accurate in numba
  • diff in numpy vs numba impl
  • numba impl may be accumulating round-off errors
• #6754 - Add the ability to alias a function/module within the compiler registry
  • Graham to respond
• #6752 - After adding nopython=True in @jit() getting error.
  • We may be able to detect user error that uses a class numba doesn't support
• #6749 - Parallel accelerator yields incorrect results for high dimensional arrays
  • has PR. small scheduler bug
• #6747 - error installing numba
• #6746 - build fails on AArch64, Fedora 33
• #6745 - Feature request: np.searchsorted(array, range)
• #6744 - LiteralSlice.indices errors

Closed Issues

• #6771 - Jitclasses and Structrefs Memory Leak Object-Like fields
• #6765 - CUDA: Cooperative groups in Windows without TCC mode
• #6764 - Numba 0.53.0rc2 checklist

2. New PRs

• #6770 - Support changing dispatcher options via a with context
• #6769 - CUDA: Replace CachedPTX and CachedCUFunction with CUDACodeLibrary functionality
• #6767 - Build script and CI cleaning
• #6766 - Fix DeviceNDArray null shape issue
• #6762 - Glue wrappers to create @overload from split typing and lowering.
• #6760 - Fix scheduler bug where it rounds to 0 divisions for a chunk.
• #6753 - The check for internal types in RewriteArrayExprs

Closed PRs

• #6763 - use an alternative constraint for the conda packages
• #6761 - Fix llvmlite dep
• #6759 - Update CHANGE_LOG for 0.53.0rc2
• #6758 - patch to compile _devicearray.cpp with c++11
• #6757 - Cherrypick patches for 0.53.0rc2
• #6756 - Cross 6750 6755
• #6755 - install llvmlite from numba/label/dev
• #6751 - Suppress typeguard warnings that affect testing.
• #6750 - Bump to llvmlite 0.37 series
• #6748 - Pr 6741 continued
• #6743 - void

3. Next Release: Version 0.54.0/0.37.0, RC=May 2021

4. Upcoming tasks