Minutes_2023_10_17

Numba Meeting: 2023-10-17

Attendees: Val, Kaustubh, Da Li, Jim Pivarski, stuart, Todd A. Anderson, Siu Kwan Lam, brandon willard, FPOC (last week): Kaustubh FPOC (incoming): Guilherme

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

Please refer to this calendar for the next meeting date.

0. Discussion

• Numba 0.58.1 and llvmlite 0.41.1 have been tagged. Artifacts available tomorrow.
  • https://github.com/numba/numba/issues/9238
  • https://github.com/numba/llvmlite/issues/999

Triage topics:

• Q. Suppose we could make it so that Numba could be imported with at least one of LLVM not initialising OR LLVM not loading into memory. Is this advantageous?
  • Notes: because TargetMachine init requires backend info and creating a depedency on backend from the frontend. Even GPU targets need to load LLVM despite not needing anything else from it.
  • Resolution:
    • worth a try as it may improve numba import time
• Given that object mode fallback is going away in 0.59.
  • a) What should we do with nopython= in all the decorators.
  • b) Should the numba code base go to using @jit exclusively and only use njit to just test that njit is still working.
  • c) Should use of nopython kwarg warn that it has no effect and will be removed in the future?
  • What to do with Flags.enable_pyobject, Flags.enable_pyobject_looplift. There's also a lot of seriously weird combinations of various fallbacks, with various liftings.
  • Material difference between not enable_pyobject and nopython?!
• For numba-as-a-toolkit, should any form of "fallback" behaviour be allowed? The pipelines are currently a list and probably should be a single thing, the C++ code also has a load of support for "falling back" to other modes.
• Does anyone want or need _GeneratedFunctionCompiler that's used by @generated_jit, if not, that can go along with the decorator.

Public meeting topics:

New "Ready for Review" PRs

1. New Issues

• numba#9237 - How to cache function which with arguments function
• numba#9238 - Numba 0.58.1 Checklist
• numba#9241 - CUDA; Need to handle incorrect CUDA_HOME more gracefully
• numba#9242 - Broken SSA for conditionally defined var
• numba#9243 - Caching causes error in structref overloading
• llvmlite#1000 - AArch64 OrcJIT segfault
• llvmlite#999 - llvmlite 0.41.1 Checklist

Closed Issues

• llvmlite#1003 - x86 OrcJIT test error

2. New PRs

• numba#9235 - CUDA: Make grid() and gridsize() use 64-bit integers
• numba#9239 - [wip] ufunc.at
• numba#9244 - Fix Recursion error in parfor lookup
• numba#9245 - Fix slow InlineInlineable
• numba#9246 - Py312

Closed PRs

• merged - numba#9236 - Fixes numba/numba#9234
• merged - numba#9240 - Cherry-picks and release notes for 0.58.1
• merged - llvmlite#996 - fix typos found by codespell
• merged - llvmlite#997 - Fix issue #880 by ensuring all sources are compiled under FreeBSD.
• merged - llvmlite#998 - adding sphinx_rtd_theme to RTD build to fix build
• merged - llvmlite#1001 - Fix / workaround for OrcJIT blocking issues
• llvmlite#1002 - Remove experimental OrcJIT support due to unforeseen issues
• merged - llvmlite#1004 - Cherry-picks 🍒 and release notes for 0.41.1

3. Short-term Roadmap

gantt: https://github.com/numba/numba/issues/8971