Minutes_2023_05_16

Numba Meeting: 2023-05-16

Attendees: Siu Kwan Lam, Guilherme, Andre Masella, Ianna Osborne, Jim Pivarski, Kaustubh, Matthew Murray, Mingjie Wang, Todd A. Anderson, vhaenel, brandon willard, Stuart, Luk FPOC (last week): Guilherme FPOC (incoming): Stuart

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

Please refer to this calendar for the next meeting date.

0. Discussion

• (val) Numba 0.57.1

  • 5 issues reported, 5 have pending PRs (at the time of writing), milestone is here https://github.com/numba/numba/milestone/71
    • ISSUE -> FIXED_BY
    • https://github.com/numba/numba/issues/8939 --> https://github.com/numba/numba/pull/8948
    • https://github.com/numba/numba/issues/8951 --> https://github.com/numba/numba/pull/8964
    • https://github.com/numba/numba/issues/8949 --> https://github.com/numba/numba/pull/8965
    • https://github.com/numba/numba/issues/8940 --> https://github.com/numba/numba/pull/8958
    • https://github.com/numba/numba/issues/8936 --> https://github.com/numba/numba/pull/8911
  • Will do an RC first
  • Aiming to get all patches merged by end of week.

• (val) llvmlite 0.40.1

  • 4 issues reported -- milestone: https://github.com/numba/llvmlite/milestone/20
    • Need to apply an SVML patch from intel, requires an llvmdev rebuild (https://github.com/numba/llvmlite/issues/943)
    • CF issue regarding PyPy builds: @siu has a patch, we need to confirm this is working and then communicate this back to CF folks : https://github.com/numba/llvmlite/issues/944 PR: https://github.com/numba/llvmlite/pull/945
    • LLVM IR generation bugs: https://github.com/numba/llvmlite/issues/946
      • Must clamp to LLVM 14 in this case, we sacrifice backwards compat.
    • now that we do need a 0.40.1 -- we can include a fix for https://github.com/numba/llvmlite/issues/941
  • Also, great opportunity to test drive: https://github.com/numba/llvmlite/pull/904 -- so we cherry-pick that one too!

• (Guilherme) Q: can llvmlite support different versions of LLVM?

  • (Andre) we need to determine the backward compatibility guarantee. LLVM IR is not backward compatible and not versioned.
  • (Guilherme) is working on a project that doesn't pin to llvm version. Endpoints may have different llvmlite version installed.
  • (Andre) If llvmlite is to become more of an independent project, we will need to have multiple IR generation backend.
  • (JimP) is working on getting info/stats on what API the users are using.

• (val) conda-forge are requesting a Numba person ask on an LLVM forum to review the SVML patch for LLVM 14 : https://github.com/conda-forge/llvmdev-feedstock/pull/192

  • Anecdotal evidence: Several attempts have been made to upstream this over the years, but none were successful. Thus it is unlikely that anyone from LLVM will come and review this patch.

• (siu) roadmap gantt as a GH issue if we agree

  • https://hackmd.io/TwFGUMjKQV6uSo0r1__KMg?both
  • decided to publish this as GH issue or as a markdown file tracked in the code base

• (collison) Can numerical fix for CUDA pow() go into 0.57.1? OK as it's fixing a numerical issue. Can llvmlite have a patch for compiler-rt added into 0.40.1? No as it's adding new features/API.

• Question about using RNG under parfors. Suggestions made to use the new NumPy RNG API which has a SeedSequence that potentially fits the usecase. xref: https://github.com/numba/numba/issues/4452. Long discussion about how to do this. Including how randint is actually parallelised. TLS based threadstates are probably impacting what happens for each chunk in the RNG. Needs further investigation, esp with respect to SeedSequence.

Updates from numba- projects? ... postponed until next time

• numba-rvsdg
• PIXIE

New "Ready for Review" PRs

1. New Issues

• numba#8950 - Type signature spelling for arguments with default values
• numba#8951 - NumbaDeprecationWarning from numba/cuda/random.py
• numba#8952 - Support for PEP 585 and PEP 604
• numba#8954 - Migrate gpuCI to RAPIDS Ops Github Actions + self-hosted runners
• numba#8956 - Weird behavior while using numba on a raspberry pi zero
• numba#8957 - Seed not working with np.random.randint only if parallel=True
• numba#8959 - A method to free the memory used by the device arrays after computations on GPU
• numba#8960 - Jupyter/Ipython Magic tools for Numba
• llvmlite#943 - SVML and SLPVectorizer patch
• llvmlite#944 - conda-forge build failures related to PyPy's ctypes
• llvmlite#946 - Wrong syntax for LLVM 14 argument attributes

Closed Issues

• numba#8955 - Unpredictable crashes with specific numba code after several consecutive runs
• numba#8961 - Unsupported .version 8.1

2. New PRs

• numba#8953 - as_numba_type: support PEPs 585, 604
• numba#8958 - resurrect the import, this time in the registry initialization
• numba#8962 - Jupyter magic command for displaying Numba IR
• numba#8963 - Fix math.pow with int64 argument for CUDA
• numba#8964 - fix missing nopython keyword in cuda random module
• numba#8965 - fix return dtype for np.angle
• llvmlite#945 - Fix #944. Add .argtypes to prevent errors in pypy.

Closed PRs