Skip to content
Valentin Haenel edited this page Oct 21, 2020 · 1 revision

Numba Meeting: 2020-10-20

Attendees: Siu, Graham, Hameer, Todd, Val, Guilherme, Stu, Juan

0. Feature Discussion

  • 0.52.0RC1 & 0.52.0RC2
    • Quickly discovered several minor issues in RC1.
    • Planned to release RC2 this week then we'll ask downstreams and users to test
    • Early feedbacks:
      • conda package issue with TBB
    • Main change in RC2 will be turning down the "pre-optimization" from O1 to O0 because LLVM is not very good at consuming it's O1 output; e.g aggressive loop-vectorization fail
  • Report benchmarks ASV
  • Upcoming tasks/risks declared for 0.53.0
    • next release planned after new year holiday
    • Hameer: working on sort algorithms; improvements and algos---timsort, radixsort
    • Siu: Getting line coverage and cProfile working. Snakeviz plot of e.g. some alg is possible (once PR merged).
    • Graham: Make the CUDA dispatcher use the same dispatcher as the CPU (the C/C++ code side of it). Some changes needed to C/C++ dispatcher which could impact all targets.
    • Stuart: Piping more information into the dispatch logic from function level to be able to do dispatch based on metadata.

1. New Issues

  • #6355 - NamedTuple type and List
    • Can probably be implemented, marking as feature request
  • #6356 - Enumerate decrefs too eagerly
    • Fixed, closed
  • #6360 - Inconsistent result for Numba in Metropolis-Hastings Distribution Implementation w/ SciPy interp2d
    • Stalled, perhaps setting the seed might help, unable to reproduce.
  • #6366 - "Numba requires at least version 0.31.0 of llvmlite."
    • waiting on answer
  • #6370 - CUDA tests: There is no RNG seeding in test_atomics.py
    • In progress
  • #6371 - 0.52.0RC1 issues
    • Fixes for most things in progress
  • #6381 - np.shift_left and np.shift_right unsupported for np.uint64
    • Tedious to implement
  • #6382 - Release global interpreter lock (GIL) in ahead-of-time-compiled code
    • Will run into the issue of config system mismatch between decorator options and Flags()
  • #6383 - Moran's I error
    • Probably fixed in newer Numba version
  • #6386 - Can't AOT String Keyed Dictionaries: undefined symbol: _numba_hashsecret_siphash_k1
    • Compiles correctly and then doesn't load
  • #6387 - NumPy ndarray expression with broadcast is slower when not use local variable.
    • Maybe fixable, if expressed as loops rather than ufuncs
  • #6388 - NamedTuple fails if type name doesn't match variable name
    • This is how Python works
  • #6389 - graphviz/dot internal buffer truncation causes issues for inspect_cfg.
    • Fix available, closed

Closed Issues

2. New PRs

  • #6353 - [WIP] Add new impl cache entry for overloads with omitted args
  • #6358 - Add log2 and remainder implementations for cuda
  • #6367 - Support views on pinned/mapped arrays
  • #6368 - Testhound/cuda atomic and
  • #6369 - disable refprune pass
  • #6372 - Skips failing tests if SVML is present.
  • #6373 - Allow Dispatchers to be cached
  • #6374 - Use Python 3 style OSError handling
  • #6375 - CUDA tests: reduce managed memory allocation to 0.1 * GPU RAM
  • #6376 - Added a fixed seed in test_atomics.py for issue #6370
  • #6377 - CUDA: Fix various issues in test suite
  • #6378 - Fix incorrect CPU override in vectorization test.
  • #6379 - Use O0 to enable inline and not affect loop-vectorization by later O3…
  • #6380 - Only test managed allocations on Linux
  • #6384 - Fix failing tests to match on platform invariant int spelling.
  • #6385 - Fix/052 rc1 octomerge1
  • #6390 - Updates inspect_cfg

Closed PRs

  • #6354 - Bump llvmlite to 0.35. series.
  • #6357 - Fix enumerate invalid decref
  • #6359 - Fixes typed list indexing on 32bit
  • #6361 - relax the llvmlite requirements for ASV and testing purposes.
  • #6362 - Move llvmlite support in README to 0.35
  • #6363 - Note that reference counted types are not permitted in set().
  • #6364 - Move deprecation schedules for 0.52
  • #6365 - Update CHANGE_LOG for 0.52.0

3. Next Release: Version 0.53.0, RC=Q1 2021

4. Upcoming tasks

Clone this wiki locally