Minutes_2020_02_18

Numba Meeting: 2020-02-18

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

0. Feature Discussion

• Issue discussion:
  • Optional<T> issues
    • #5259, #5258, #5200
    • Stuart asks if we can remove Optional type
      • kill off branches on optional
      • refine types
      • try and minimise their creation
      • Dict.get() can still return None dynamically though
      • need help thinking of a good solution
  • Nested parallel region behaviour
    • single-function nested parallel region is still sequential-ized by the compiler
    • cross function parallel region can launch new parallel tasks and cause throttling
    • currently user control with set_num_threads()
    • may become a problem
  • chainsawing status
    • del removal
      • goal: no ir.Del in IR until lowering
      • background: Dels are used as markers for lifetime
    • parfor code refactoring and finer-grain tests
  • function 1st class type status
    • making progress
    • Pearu will remove the sig arg in WAP and we can review+merge https://github.com/numba/numba/pull/4967
    • njit support will come in another PR.

1. New issues

• #5264 - Heap corruption of Lists of numpy arrays
  • probably related to TypedDict.get https://github.com/numba/numba/issues/5081
  • probably decref before incref
• #5263 - parfors argmin on boolean array fails
  • easy fix
• #5261 - Can we define char array by Numba CUDA?
• **** related
  • #5259 - Raising an error casts result to OptionalType
  • #5258 - OptionalType pruning after test for None
• #5256 - Moved numba.targets causes AttributeError for pandas
  • we'll add shim and the is_jitted() check
• #5255 - .dtype cannot be in the jit?
• #5254 - forceobj=True and loop with continiue
• #5251 - Test failures in numba master
• #5248 - double free or corruption (!prev) Segmentation fault (core dumped)
  • may want to make List going in ParFor loop immutable
• #5246 - Bitwidth of datatypes should be an immutable property or name should change on assignment
• #5245 - Unable to create record array
• #5244 - myarr.searchsorted(...) raises TypingError while np.searchsorted(myarr, ...) works
  • wiring problem
• #5243 - Tuple in np.array constructor fails without clear error only upon direct assignment to secondary array
  • SSA
• #5242 - literally() fails when used in implementations with default arguments and overload_method
• #5241 - Typed list construct from iterable.
• **** #5236 - Unboxer/boxer caching for speed
• #5234 - Numba DeviceNDArray doesn't handle shape=() correctly
• #5233 - Add dtype property to List class
  • associated PR#5235
• #5229 - Calling .ravel() on a DeviceNDArray errors

Already Closed

• #5260 - array(bool, ...) are not isinstance(..., types.Boolean)
• #5238 - Test discovery with vscode fails
• #5228 - Parallel prange (still) gives LoweringError

2. New Open PRs

• #5265 - Move preparfors pass
• #5262 - fix typo in inlining docs
• #5257 - boundscheck fix
• #5253 - Coerce dtype to numpy.dtype
• #5252 - Trivially fix SyntaxWarning
• #5250 - Part fix/5098
• #5249 - fix target path for See also
• #5247 - Continue PR #5126
• #5240 - Help needed: test failures with overload
• #5237 - Fixes Parfor nested-loop transformation with Python 3.8
• #5235 - Added property dtype to numba.typed.List
• #5231 - [WIP] added support for np.asarray to be used with numba.typed.List
• #5230 - Show bitwidth in string representation of IntegerLiteral
• #5227 - Immutable typed-list

Already merged/closed

• #5239 - fix test discovery for unittest
• #5232 - Fixes a bad import that breaks master

4. Next Release: Version 0.49.0, RC=March 19

• Requests for 0.49
  • TBD