Minutes_2021_09_28

Numba Meeting: 2021-09-28

Attendees: Siu Kwan Lam, Graham Markall, Jim Pivarski, Nick Riasanovsky, stuart, Valtentin Haenel, Luk, Guilherme Leobas, Milind Chaudhari, Shi Fan, Todd Anderson

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

0. Feature Discussion

• 0.54.1rc1
  • In RC testing phase.
  • we have to patch vendored cloudpickle w.r.t. dynamica class unpickling. Folks using distributed frameworks and caching extensively are encouraged to test this RC.
• Stuart is experimenting with pexpect to interact with gdb.
  • related PR#7177
• https://github.com/numba/numba/pull/5209 - importlib
  • importlib API changes across python versions and importlib versions
• python 3.10
  • with-statement support affected by bytecode inliner
  • we need to reduce the reliance on heuristics on bytecode sequencing by moving to better tracing and controlflow analysis.
• Discussion on https://github.com/numba/numba/issues/7436
  • related test: https://github.com/numba/numba/blob/e617b39a0b4b23d7b69d16f482fd66b4ac6cc307/numba/tests/test_parallel_backend.py#L867-L956
• Chrome event trace PR is WIP
• Reminder for extension writer:
  • https://github.com/numba/numba/pull/7397 will change the error handling behavior.
  • There will be more "hard" errors---no fallback.
  • "new_style" errors, or "hard" errors, will become a default in a few releases.
• Q: How to profile compilation in a section of a program?
  • A: Event API: https://nbviewer.jupyter.org/github/numba/numba-examples/blob/master/notebooks/Numba_053_profiling_the_compiler.ipynb

1. New Issues

• #7412 - Strange bug when combining jitclass, type annotation and TypedList
  • Is this an issue with the __annotations__ that can not be serialized?
  • Do we need serializers for annotated classes?
  • Can this be solved with a MetaKlass?
• #7413 - Default constructor of jitclass not useable from jitted code
  • Dup: https://github.com/numba/numba/issues/4495
  • No kwargs in Jitclass
• #7414 - Show function arguments does not work for debugging w/ GDB
• #7415 - Function argument are not available at the first line when debugging w/ GDB
• #7416 - Using literal dicts fails to compile when string keys hold numeric data
  • Maybe doable with a StructRef
• #7417 - [Feature Request] Support for Python collections.deque+
  • Pronounced "deck"
  • Maybe a computer science 101
• #7420 - Debugger incorrectly shows zero value for variables
• #7423 - numpy 1.21 ufunc reduce segfault
• #7424 - numba.jit strips away function type annotations, breaking mypy's analysis
• #7426 - Loss of functionality in .54 - change in slicing behavior on user defined types
• #7427 - iter(Dict()) shows odd behavior + iter(List()) breaks
• #7431 - Numba 0.54.1rc1 Checklist
• #7434 - TypeError when calling jitclass method with optional type member
  • Error might be weird, it's doing the right thing.

Closed Issues

• #7418 - Warnings regarding ROC toolchains, libhsa-runtime64.so
• #7433 - dump-llvm hangs

2. New PRs

• #7421 - Add option to extend variable lifetimes to the end of their block.
• #7422 - Update Omitted Type to use Hashable Values as Keys for Caching
• #7425 - fix: preserve decorated function's type annotations
• #7432 - add github release task to checklist
• #7435 - Make branch pruner pass SSA compliant

Closed PRs

• #7419 - insert missing backtick in comment
• #7428 - Fix regression #7355: cannot set items in structured array data types
• #7429 - Update CHANGE_LOG for 0.54.1
• #7430 - Cherrypicks for release 0.54.1

3. Next Release: Version 0.55.0/0.38.0, RC=Early Nov

• Request for 0.55
  • broadcast_to https://github.com/numba/numba/pull/7119
    • driven by need for vectorize support in aesara
    • Now merged