Minutes_2023_01_24

Numba Meeting: 2023-01-24

Attendees: Val, Graham Markall, Kaustubh Chaudhari, stuart, Todd A. Anderson, Guilherme, Andre FPOC (last week): Stuart 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

• General updates

  • Python 3.11, pretty much done
    • Missing 8 tests
  • NumPy 1.24, pretty much done, couple of things to resolve
    • Waiting for packaging and minor things to resolve
  • LLVM 14, public CI is passing, working on platform specific issues on the build farm.
  • ORC JIT 2 - now does everything that MCJIT does except object caching and add object file. Works on RISCV too (with a few fixes backported)! Should be able to try with Numba soon.
    • try and get this in for 0.40 depending on review load.
  • objmode fall back is now officially going to be removed in Numba 0.59.
  • There are a few critical issues to resolve for 0.57, most of which are to do with deprecation notices or updates to core packages.

• Guilherme: Question about new frontend work... status update:

  • Val: it's being worked on, hoping to present progress within the next couple of months.

• Todd: Question about caching: If you dynamically create a function, and then njit(cache=True) that function. It appears as a <string> defined function, can this be cached?

  • (stuart) It depends on the application. Falsify a location for the generated function. But how does one identify that two functions are the same, with globals/locals and free variables?
  • In limited cases, the <ipython- cache might be sufficient if replacing <string> with other piece of knowledge.
  • Ideally fix (#6406) which attempts to make caching of string defined functions work.
  • luk-f-a is working on a patch that contains a lot of test cases for cache invalidation (#8369)

New "Ready for Review" PRs

• https://github.com/numba/numba/pull/8396 - Extend cache invalidation to include changes in other files (take 2)

1. New Issues

• numba#8717 - compile_ptx() allows compilation of kernel with non-void return type
• numba#8718 - Numba __init__.py not catching incompatible NumPy version early enough
  • stuart will write patch to fix, @gmarkall will review.
• numba#8719 - CUDA simulator does not raise for more than one group in cooperative groups mode
• numba#8722 - problem with atomic.compare_and_swap
• numba#8724 - save and read numba.typed.Dict , with multiple workers
  • Would be good to speed up typed container constructors.
• llvmlite#900 - llvmlite opaque pointer support
  • Wait until LLVM 14 patch is merged, then start this work.

Closed Issues

2. New PRs

• numba#8720 - Enable coverage for subprocess testing
• numba#8723 - Check for void return type in cuda.compile_ptx

Closed PRs

• numba#8721 - [CI only] Release 0.56 + CUDA lineinfo fixes

(last numba: 8724; llvmlite 900)

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