Minutes_2020_07_21

Numba Meeting: 2020-07-21

Attendees: Siu, Graham, Guilherme, Todd, Ehsan, Stuart, Val,

0. Feature Discussion

• RC1 postpone? Aug 5?
  • llvm 10 issues
    • aarch64 copy reg bug
    • armv7l unknown large memory use
  • stabilizing master
    • overspecialization issue
      • switch around--don't prioritize literal
  • loads of tbb problems
• A discussion on Bodo's system
  • a closed system
  • tries to provide better errmesg @overload
  • Ehsan to open a PR to contribute their flag for literal
  • Siu talks about a exception formatting callback
• mypy on CI https://github.com/numba/numba/pull/5960
  • Ehsan thinks this will improve code quality
  • Siu think so too. Only slight concern on readability.
  • Siu this will help llvmlite builder code
  • Stuart at +0.5. investigate compound type
  • to be merged for this release
• Semantics for cuda.atomic.nan{min,max} returned value (PR 5941) - when old value is NaN, return:
  • Old value?
  • New value?
  • Have no return value?
  • preferred old value for simplicity in description
  • need to check datashader issue; old code is different
  • Graham to take over the PR
• Ehsan looking at contributing on caching
• Ehsan discussing OrcJIT vs MCJIT
  • lazy compilation w/ OrcJIT
  • functions that are opt-away is being compiled by eager MCJIT

• Guilhereme on https://github.com/numba/numba/pull/5938

1. New Issues

• #6012 - BUG: Symbol not found: __powidf2 on Win32
  • New LLVM problem that didn't happen before LLVM9
• #6011 - Signal handling inside compiled code
  • shouldn't be too hard
• ** #6009 - Extremely long compilation time for a tiny function
  • another problem with long tuple
  • maybe it's time change the calling-conv to pass large immutable struct/tuple by reference like C++ ABI
  • make boxing/unboxing reuse function
• #6007 - Type inference failing on code explicitly not ran with certain variable type
• #6005 - Error hidden when source is not available
• #6003 - Release 0.51.0RC1 / 0.34.0RC1 Checklist
• #6001 - typed.List ignores constructor arguments when jitting is disabled
• #6000 - Updating 1d array with 0d array fails
• #5998 - numba installation on nvidia Jetson
• #5997 - Self-referential types cannot be unpickled
• #5994 - Literal sets missing implementations
• #5992 - Misleading Documentation for synchronous CUDA Kernel
• #5991 - Reminder to update CONTRIBUTING.md and add "good first review" label and use it
• #5990 - differentiate between cpython and numpy scalars
• #5989 - Update tests use of Popen
• #5988 - Add utilities to the namemangler to produce things like "the Numba jit/cpython wrapper prefix"
• #5987 - Passing objects as input of jit.cuda function
• #5985 - NumPy floor_divide zero handling wrong value produced.
• #5983 - After computing delayed dask from numba, opening subprocess hangs on repeated sched_yield
  • horrible problem with TBB x fork
• #5982 - Intel TBB is required, no threading layer could be loaded, however TBB is already installed

Closed Issues

• #6002 - Global Random Seed
• #5999 - copy() or ascontiguous() functions takes considerable amount of time in python or numba, especially if it is a big array.

2. New PRs

• #6010 - Set Literals
• #6008 - Small fix so bullet points are rendered by sphinx
• #6006 - Resolves incorrectly hoisted list in parfor.
• #6004 - Fix problem in branch pruning with LiteralStrKeyDict
• #5986 - flake8: numba/core/annotations/*.py
• #5984 - Handle idx and size unification in wrap_index manually.
• ** #5980 - Deprecate target kwarg for numba.jit.
  • decide to deprecate in 2 releases.
  • TODO:
    • update docs and warning msg.

Closed PRs

• #5996 - Set lower bound for llvmlite to 0.33
• #5995 - Update setup.py with new llvmlite versions
• #5993 - Docs for structref
• #5981 - Document bounds-checking behavior in python deviations page

3. Next Release: Version 0.51.0, RC=22 July, Final 29 July?

• Requests for 0.51

• high risk stuff for 0.51.

• 0.51 potential tasks (To be updated)

4. Upcoming tasks