Skip to content
Valentin Haenel edited this page Aug 25, 2021 · 1 revision

Numba Meeting: 2021-08-03

Attendees: Siu Kwan Lam, Todd Anderson, brandon willard, Graham Markall, Lehman Garrison, stuart, Jim Pivarski, Luk

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

0. Feature Discussion

  • PSA:
  • Exposing control over optimization in pipeline: req discussed in Discourse
    • Q: How to expose fine-grain control over optimization? (related post)
    • Ideas:
      • Allow user to pass in MPM?
      • 0-10 optimisation level arg for functions
      • Suggest user creates their own higher-optimization effort JIT
      • Or start at aggressively optimize and let user turn it down.
      • Genetic Algorithm for pass selection (Stuart has a prototype)
        • is observed to sometimes outperform O3
    • Side Questions:
      • Q: Why multiple passes of optimizations; e.g. "cheap", then "refct", then "expensive"?
      • A: We have "refct" pass to remove atomic reference count operations, which can block aggressive loop optimizations in "expensive" pass. The "refct" pass does best when the IR is inlined and simplified by the "cheap" pass.
  • Perf hack: https://gist.github.com/sklam/30c2ff32ac72be197a349e6c4e1f5d32

1. New Issues

  • #7264 - StrEnum comparison is incorrect inside jit or njit wrapped functions
  • #7261 - Update URL in error message to use issue templates
  • #7259 - More efficient np.zeros
  • #7258 - Reassign to an IR Variable Produces a Memory Leak
    • Siu to check
  • #7254 - Unused closures in njitted function are wrongly reported as escaping
  • #7253 - CUDA: Multi-GPU tests failing on Windows
  • #7250 - Typed List indices are cast unsafely

Closed Issues

2. New PRs

  • #7263 - Updated issue URL for error messages #7261
  • #7262 - fix order of handling and casting
  • #7260 - Support typed lists in heapq
  • #7257 - Use normal list constructor in List.new()
  • #7256 - Fix link to the discourse forum in README
  • #7255 - CUDA: Support CUDA Toolkit conda packages from NVIDIA
  • #7252 - Add numpy.delete optional argument axis support

Closed PRs

  • #7251 - Fix #6001: typed.List ignores ctor arguments with JIT disabled

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

Clone this wiki locally