Minutes_2020_01_28

Numba Meeting: 2020-01-28

Attendees: Aaron, Graham, Stuart, Todd, Val, Pearu, Siu

0. Feature Discussion

• 0.48.0 release status

  • tagged and building
  • upload by EOD

• __cuda_array_interface__ https://github.com/numba/numba/issues/4933#issuecomment-572934894

  • Folks seems to agree an update to docs: consumer must use legacy default stream
  • double-check on the cuda docs on the implicit synchronization and mention non-blocking stream

• Chainsawing

  • remove py2.7, py3.5, old numpy<1.15
  • removed numba.six replace it with numba.utils
  • single spelling for python version and numpy version

• File moving plan

  • Plan: https://hackmd.io/XaB_9Qq2SnWYxhUvEm6pEw
  • Git hints: https://stackoverflow.com/questions/3491270/git-merge-apply-changes-to-code-that-moved-to-a-different-file
  • no file splitting yet

• TODOs from last meeting:

  • Share plan of chainsawing
    • will be investigating the impact of using black
  • TODOs for others orthogonal to chainsaw (once file move done, @overload replacement)

• Issue discussion:

  • #4809 Warnings executing numba random number generator example with simulator enabled
  • NBEP 7 - planning to go ahead with implementing as described in doc

1. New issues

• #5145 - Checklist for 0.48.0 release
  • task
• #5143 - TypingError with overload(..., inline='always') and generated implementation used in a closure
  • not sure yet
• #5142 - inline parameter not honored when function is called in an overload implementation
  • need to ask for usecase
• #5140 - Switch Azure CI windows config to use VS2017 images
  • task
• #5139 - Numba can't vectorize simple loop
  • 1. needs new llvm fn attributes
  • 2. call to a fastmath can miss the "fast" attr (need to investigate)
• #5138 - Running python3 script results in numba error
  • llvmlite version issue
• **** #5137 - Support per-thread default streams in CUDA
  • suggest to look at localized per-thread config.
• #5135 - Support for dictionary comprehensions.
  • Need better err msg
  • can probably support limited cases of this now
• #5132 - Error with umap.plot.diagnostic
  • prange tuple problem
• #5131 - CUDA: pinned_array_like, pinned_zeroes, pinned_ones, etc. would be handy
  • yes
• #5129 - Key Error in numba annotations when caching is turned on
  • confirmed bug
• ** #5128 - make caching work for source-less functions, e.g. strings, lambdify exprs.
  • maybe we can stop relying on files for modification timestamp
  • what if we just rely on the code object and hash of globals
    • but non-hashable globals?
  • need more thoughts
  • TODO: put notes the issue(s)
• #5125 - Passing a set of strings to Numba function fails
  • fix "Hashable" of str
  • set is set to be rewritten
  • to fix the call conversion problem
• #5124 - slicing arrays pessimises array layout in typing
  • perf problem
  • potential enhacnement: static_getitem pass can strength reduce
• #5121 - Issue with NUMBA
  • insufficient info
• #5120 - numba.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
  • closed
• #5116 - ord builtin not supported
  • PR for ord/chr opened https://github.com/numba/numba/pull/5117

Already Closed

• #5141 - inline=
• #5134 - Numba missing dependency for setuptools
• #5133 - Issues with operator.is(int, int)
• #5130 - CUDA device to host transfer with a stream synchronizes with device
• #5109 - Numba nested function won't execute (version 0.47)
• #5107 - Imprecise result when raising complex128 to power of complex128
• #5096 - Compilation time of a njit function getting too long when one parameter is a large list of jitclass
• #5095 - In jitclass, will not compile when a member variable is used in numpy.zeros()
• #5089 - Numba 0.48.0rc1 Checklist
• #5087 - Function working with 0.46 doesn't work with 0.47 (runs indefinitely)
• #5082 - Typing error with overload of function without argument and without return
• #5080 - dtype=np.bool doesn't work in array constructors
• #5077 - Please include support for Python exception handling
• #5076 - numba.errors.LoweringError: Failed in nopython mode pipeline (step: nopython mode backend)
• #5070 - is there any Thread Safe Container ?

2. New Open PRs

• #5147 - Update master for 0.48.0 updates
• #5144 - Fix #4875: Make #2655 test with debug expect to pass
• #5136 - CUDA: Enable asynchronous operations on the default stream
• ** #5127 - Calling convention adaptor for boxer/unboxer to call jitcode
• #5126 - Modules can't be parfor races.
• **** #5123 - [DISCUSS] Run CUDA tests on CUDA hardware with gpuCI
• ** #5122 - Remove Python 2.
  • six -> numba.utils
  • no more weird spelling of python versions
• ** #5119 - Make IR nodes immutable.
  • keep around for reference
  • can't easily merge
• #5118 - Fix readability issues in the docs
• #5117 - Implement ord()/chr()
• #5115 - Add support for iterating over 2D arrays
• #5114 - Correctly restore numba.parfor.sequential_parfor_lowering in case of …
• #5113 - Fix error handling in Interval example
  • may want to put the ExitStack trick into pyapi.py instead

Already merged/closed

• #5146 - Update 0.48 branch
• #5106 - Add SciPy 2019 talks to docs
• #5092 - Fix #5087 bug in bytecode analysis
• #5091 - Change log update for 0.48.0
• #5090 - Update deprecation notices for 0.48.0

4. Next Release: Version 0.49.0, RC=March ???

• Requests for 0.49
  • TBD