Minutes_2023_03_14

Numba Meeting: 2023-03-14

Attendees: Jim Pivarski, Graham Markall, collison, Andre Masella, Ianna Osborne, Brandon Willard, Val, Stuart, Kaustubh, Guilherme, Siu Kwan Lam, Todd A. Anderson FPOC (last week): Andre FPOC (incoming): Graham

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

Please refer to this calendar for the next meeting date.

0. Discussion

• 0.57 release prep status:
  • Numpy 1.24 (merged?)
  • LLVM14 merged
  • Py3.11 merge preparation in progress
• LLVM 14 AArch64 blocker
  • Numba Issue #8738: LLVM14 linux-aarch64 blocker
  • LLVM Issue: #61402: RuntimeDyldELF doesn't clear GOTOffsetMap in finalizeLoad(), leading to invalid GOT relocations on AArch64
  • Patch: https://github.com/numba/numba/issues/8738#issuecomment-1467136997 - with this patch:
    • All LLVM unit + regression tests pass on AArch64 + x86
    • All Numba tests pass with Numba main and llvmlite main
  • How to proceed?
    • This is one more patch on the pile, so no change to status quo
    • Graham to make PR to add patch to llvmdev build
    • Andre to run through buildfarm
    • Ping Debian etc. maintainers
• Should SLP be turned off again? [stuart]
  • Seems to cause failures to correctly initialize registers, involved in:
    • #8705: np.cos, np.sin, np.sqrt returning incorrect values in very specific scenarios
    • #8681: Numba crashes Python with simple function containing numpy.exp
  • Turning it off may have some performance impact, but seems to be causing some miscompiles
    • Use a/the flag to turn it on?
    • Consensus steering towards correctness over performance
• [name=JimP]: Invitation to PyHEP.dev for coordination between Numba development and HEP (High-Energy Physics) software development. Contact pivarski@princeton.edu.
• Can support for compiler-rt be added to the next version of llvmlite (https://github.com/numba/llvmlite/pull/909)?
  • llvmlite release cycle independent of being able to use this PR once merged.
  • Once it is merged, we'll rebuild llvmdev and it will be usable for Numba CI from then on.
  • Potential workarounds for the conda-build issue on macOS.
    • https://github.com/conda/conda-build/issues/4787
  • Will this enable arbitrary precision integers?
    • Possibly, would need to check what compiler-rt implements
• Compiler plugins support in llvmlite: https://github.com/numba/llvmlite/pull/915
  • Shared lib build issues
  • llvmlite on numba channel statically linked, but dynlinked elsewhere
    • also statically linked on wheels
  • what do we really need to do to fit the goal?
    • in support of enzyme for numba
  • other option to provide a way to build passes into llvmlite
  • option: implement as is in pr but you need to link against llvm dynamically and we cant guarantee it works.
  • another (maybe not attractive) option: llvmdev package wheel with so and a small import "stub"
  • Symbol versioning also an issue
    • related issues: https://github.com/llvm/llvm-project/issues/47565
  • Possibly another option better: serialise to bitcode and run through pass externally then pipe back in

New "Ready for Review" PRs

1. New Issues

• numba#8797 - Allow typed containers to be pickled
• numba#8798 - Support python sortedcontainers in nopython mode
  • close. out of scope for numba.
• numba#8799 - ImportError: cannot import name 'njit' from 'numba
• numba#8800 - double free malloc error in generator returning reference
• numba#8801 - CUDA: Error when linking C/C++ sources when Minor Version Compatibility is enabled
• numba#8806 - Bug in numba recipe: zipping tuples.
• numba#8808 - jitclass doesn't work when dict class attribute has array as value type
• numba#8812 - <intrinsic to_fixed_tuple> is not usable in pure-python
• numba#8813 - LLVM14 SVML causes small change in precision in test_logaddexp2_dd_d on osx
  • up the ULP to fix
• llvmlite#912 - “Escape hatch” for maximum Python version check
  • remove max python version guard
• llvmlite#913 - LLVM14 test suite not run on llvmdev pacakage build
• llvmlite#914 - Support for LLVM 15.0.7
• llvmlite#916 - cannot install llvmlite via poetry

Closed Issues

• numba#8805 - 'numpy.dtype[float64]' object has no attribute 'is_precise'

2. New PRs

• numba#8802 - Move the minimum supported NumPy version to 1.21
• numba#8803 - Attempted fix to #8789 by changing compile_ptx to accept a signature instead of argument tuple
• numba#8804 - Split parfor pass into 3 parts
• numba#8809 - Update LLVM versions for 0.57 release
• numba#8811 - Temp/debug8134
• llvmlite#915 - Add Support for Compiler Plugins

Closed PRs

• numba#8807 - [CI] Pr/8535 LLVM14 testing
• merged - numba#8810 - Fix llvmlite dependency in meta.yaml

3. Next Release: Version 0.57.0/0.40.0, RC Jan 2023