Minutes_2019_11_12

Numba Meeting: 2019-11-12

Attendees: Siu, Stuart, Todd, Aaron, Stan, James

0. Feature Discussion

• Python 3.8: The Eternal Quest

  • Trying to share more code paths between 3.8 and pre-3.8.
  • nice side effect is better analysis of complex control flow (try/except)
  • loop-lifting is the last blocker

• Pandas taking Numba as optional dependency by end of year

  • Pandas already dropped Python 2

• Thread controls

  • TLS inheritance issue
    • resolution for max safety:
      • default to 0, which means 1
  • Need to disable parallel gufunc if threadcount==1 in nested-parallel-function

1. New issues

• #4816 - Some (but not all) unicode characters causing errors in jitclass
  • ϕ vs φ why!?!?!??!
• #4814 - structs not supported for jitclass
  • not supported yet
• #4812 - Error in using ncpdq
  • probably fixed if upgrade numba
• #4810 - Performance improvements needed for transpose
  • reminder for us
• #4809 - Warnings executing numba random number generator example with simulator enabled
  • stan needs to investigate
• **** #4808 - Numba future feature flag design
  • suggestion to make the flag callable to make linters happy
• #4807 - Cache causes Segmentation Faults when generated in parallel
  • stuart has a fix
    • a read before write problem
    • index-write should happen after data-write
• **** #4806 - Invalid result with parfor
  • unsupported usage for now
  • a recent PR would flag this
  • Todd to look at this more
  • see PR#4803 to flag this issue
• #4805 - Support np.roll's Axis argument
  • axis for roll very different than other numpy usage of axis
• #4800 - AttributeError when allocating "list" in device if NUMBA_ENABLE_CUDASIM=1
  • need to asarray lists when running on the simulator
• #4799 - make html does a full rebuild every time
• **** #4797 - CUDA API interoperability
• #4795 - Does numba support sympy?
  • Answered, closing
• **** #4793 - Vectorize function (with float64 argument) raising unexpected RuntimeWarning
  • strangest bug all year
• #4790 - Refactor Task. Flags mapping to jit options
• #4788 - Performance tracking for hnswlib
• #4786 - Return types for bit shift operators
  • python / numpy integer typing 🥫of 🐛
• #4784 - Enable more efficient autovectorisation (either through jitclasses or with immutability attributes)
• #4783 - OSError: exception: access violation reading ... in nvvm.py
  • stan will respond

Already Closed

• #4801 - TypingError with random.random() if not NUMBA_ENABLE_CUDASIM=1
• #4792 - Issue Report - New to bug reporting onGitHub

2. New Open PRs

• #4815 - [WIP] Add support for numpy.isin
• **** #4813 - [WIP] implement a specialized map() for mixed-type tuple
• **** #4803 - Better compiler error messages for improperly used reduction variables.
• #4798 - Add branch pruning based on raw predicates.
• #4785 - Treat 0d arrays like scalars.

Already merged/closed

• #4796 - Try out conda binaries
• #4804 - [DO NOT MERGE] Test CI with SVML Fix
• #4811 - fix spelling in now-failing tests
• #4802 - put git hash into build string
• #4794 - Add setuptools as a dependency
• #4791 - fix typos
• #4789 - fix typos in numba/targets/base.py
• #4787 - fix missing incref on flags
• #4782 - Fix issue with SVML (and knock-on function resolution effects).
• #4781 - WIP/runtest cf pytest

4. Next Release: Version 0.47.0, RC=December 19

• CPython 3.8

• Requests for 0.47 (last release for the year) - jitclass performance issues - llvm 9 trial - CTK libcudadevrt.a - CI needs to take 50% of current time - Val & Stu already looking at this - also checking Azure CI config to avoid wasting compute time - Caching: - transitive dependency - other issues: i.e. function as argument, with objmode - distributing cache - Immutable list and deprecating reflected list - Switch to pytest (see above) - Using Numba to generate LLVM/NVVM IR for different targets https://github.com/numba/numba/issues/4546 - @overload for gpu