Minutes_2022_01_18

Numba Meeting: 2022-01-18

Attendees: Siu Kwan Lam, Graham Markall, Nick Riasanovsky, Todd A. Anderson, Ben Graham, stuart, brandon willard, Val, Guilherme Leobas, Jim Pivarski

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

Please refer to this calendar for the next meeting date.

0. Feature Discussion

• 0.55.0 released quickly to address numpy CVE affecting np<1.21.
• 0.55.1 bug fix release
  • https://github.com/numba/numba/milestone/62
  • Plan for release Thursday 27th Jan
• Release dates / roadmap
  • e.g. WRT CPython, NumPy, etc.
  • CPython 3.11:
    • Bootstrap entire stack with numba-hatchery to work with CPython 3.11 on buildfarm
    • To complete most PRs or know what issues are close to the 3.11 release
    • Aiming for anaconda.org 3.11 interpreter out soon after release
    • CPython 3.11 - ETA 3 October 2022: expect to support in Numba 0.57
  • NumPy 1.22
    • Support in Numba 0.56 (next release)
  • NumPy RC phases:
    • Investigate testing Numba earlier with RCs
    • RC period 1 month
    • NumPy release schedule for 1.23 maybe 6 months? (need to check) - support in Numba 0.57?
  • LLVM: relatively stable
    • llvmlite compatibility with multiple LLVMs
    • wait for 13.0.1 for llvmlite
    • Support in 0.56 or 0.57 (unknown)
• 0.56 Aims / goals
  • Release ETA: May burn-down
  • CUDA dispatcher to be closer to core dispatcher
  • importlib_metadata #5209
  • master -> main rename (llvmlite + Numba)
    • Do right now because of release branch
  • Iteration over empty containers (Guilherme Leobas)
  • Caching with passed functions broken because of first-class function types (bug fix)
  • Performance regression with iteration on any sort of array
  • Conversions to overloads (Stuart)
    • Fixing flags may be needed
  • Added multiple axes support for np.expand_dims #7604
  • Performance for reduction loops (Brandon Willard)
  • PR #6939 - boxing for slices - waiting for comment from reviewer
  • Updates to Array Analysis (upcoming PR for next week) (Nick Riasanovsky)
  • Update sets to have support for refcount objects (early next week)
    • Noting set bugs #3772, #6543
    • May be quicker to implement typed set

1. New Issues

• #7719 - The ability to set fma, prec_div, and prec_sqrt options dynamically for cuda.jit
• #7723 - Explicitly catch writing to globals and document?
• #7725 - Raise an error when an array of the wrong types are used for function compiled AOT
• #7726 - DWARF generation failing for OmittedArgs
• #7727 - improve performance of numba.typed.List constructor with Python list as arg
• #7730 - "Unexpected" indentation can break DI emission
• #7734 - memcpy_region looks suspiciously wrong
• #7735 - Numba 0.55.0 FINAL Checklist
• #7739 - Refactor CUDA target module and conceptual layout to match CPU
• #7741 - Refactor the CUDA target Dispatcher class to inherit from numba.core.dispatcher::Dispatcher
• #7742 - Reduce the CUDA target launch overhead by JIT compiling the launch code
• #7743 - Make it possible to launch CUDA kernels from CPU JIT code regions.

Closed Issues

• #7720 - The ability to set nogil and parallel options for cc.export
• #7721 - SortedDicts()
• #7731 - Accelerate release of 0.55 to no longer depend on versions of numpy with buffer overflow
• #7738 - 0.55 regression around (?) setitem(readonly array(int64, 1d, C), Literalint, Literalint))
• #7746 - TestThreadingLayerPriority tests fail on openSUSE rpm build
• #7747 - numba==0.55.0 is broken with numpy==1.19.5 on Python 3.7

2. New PRs

• *** #7717 - [WIP] Keep references to dynamic global arrays
  • More restrictions on globals added in 0.55 (writing to readonly, etc)
  • CPython approach (reconstructing based on module path) vs holding refs
• #7722 - Testhound/fp16 divide
• #7724 - Update URLs in error messages to refer to RTD docs.
• #7728 - Document that AOT-compiled functions do not check arg types
• #7729 - Handle Omitted/OmittedArgDataModel in DI generation.
• #7733 - fix for /tmp/tmp access issues
• #7737 - Tidy up memmove/memcpy APIs
• #7740 - [WIP] CUDA Python 11.6 support
• #7744 - Fix issues with locating/parsing source during DebugInfo emission.
• #7745 - Fix the release year for Numba 0.55 change log entry.
• #7748 - Fix #7713: Ensure _prng_random_hash return has correct bitwidth
• #7749 - Refactor threading layer priority tests to not use stdout/stderr

Closed PRs

• #7718 - Bump llvmlite dependency to 0.39.0dev0 for Numba 0.56.0dev0
• #7732 - update release checklist following 0.55.0 RC1
• #7736 - Update CHANGE_LOG for 0.55.0 final.

3. Next Release: Version 0.55.0/0.38.0, Final