Minutes_2020_02_24

Numba Meeting: 2020-02-24

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

0. Feature Discussion

• Stuart looking at "public" API
  • https://github.com/numba/numba/issues/5214
  • stable extension API
  • unstable but available API
  • a iterative process to "grow" this public API set

1. New Issues

• #5304 - Function calls cannot combine tuple unpacking and keyword arguments
  • feature request for bytecode we don't support
• #5303 - Call to cuInit results in UNKNOWN_CUDA_ERROR
  • cuda
• #5302 - datetime instances cannot be cast to type datetime
  • probably cast not implemented
• #5301 - How to include LAPACK functions with complex inputs
  • a question. answered
• #5300 - Partial tuple unpacking fails to compile
  • UNPACK_EX bytecode not supported
  • tricky; depends on compile-time knowledge of the tuple size
• #5298 - inline="always" flag for @overload_attribute
  • CLOSED
• #5291 - Compilation failure when overload contains *args and multiple keyword arguments
  • variation of #5304
  • Opened PR #5292
• **** #5290 - prange fails to generate reproducible random sequence with printing
  • (likely) Timing issue
  • or not; it maybe operating by design---randomly init in each thread
• #5289 - typed-list fails to refine via extend on unicode input
  • PR 5295
  • Need a better pattern to do this.
• #5286 - ImportError: Numba could not be imported
• #5285 - Return Numba types from @njit functions
  • bug; need JIT-transparency
• **** #5284 - Extension types loaded via entry points are registered too late
  • Possible options:
    • Lazy opt-in load, basically make it user configurable
    • Some sort of convention based on e.g. file name to force early load
• **** #5282 - Numba (parallel=True) does not handle tuples in certain situations
  • Related PRs: #5283, #5299
  • Investigate if there's a limit in LLVM etc
  • Merge PRs, Todd to pull in patch.
• #5279 - parallel=True and if-condition
  • bug. likely trouble in array_analysis.py
• #5275 - Threading backend using omp_set_nested which is deprecated in OpenMP 5?
  • depreacated in anaconda-shipped MKL 2020
• #5274 - Attempting to iterate NoneType instance leads to infinite loop
  • bug
• #5269 - Unable to pickle DeviceNDArray
  • needs to make better error
  • don't know what to do at unpickling
• #5267 - byte code analysis problem with while(True) (but legal?!)
  • byteflow bug?
• #5266 - Numba documentation page does not provide different versions
  • doc issue
  • good first

Closed Issues

• #5288 - Numba/numpy report different results for SUM(x - MEAN(x))
• #5277 - LoweringError when passing in prange values as a tuple
• #5273 - Jitting changes result when adding certain kind of 0 to local variable

2. New PRs

• #5299 - Initial attempt to support tuples to parfors.
• #5297 - Update docs with intro to {g,}ufuncs.
• #5296 - Refactor parfor: better exception from passes
• #5295 - refine typed-list on unicode input to extend
  • Need better pattern for detecting iterables.
• #5294 - Create typed-list from iterable
• #5293 - [WIP] np.isnat
• #5292 - fix argument binding with var positional and kwargs
• **** #5287 - First-class function type [WIP]
  • Can these be the default?
    • Maybe not as it's not always possible to represent a function as a CFUNC.
  • What's the impact of this on downstream projects like SDC and also on Parfors internally, will this break transforms etc?
  • Is it possible to do this dynamically, as a function arrives as an arg?
• #5283 - Improve error location and reporting for tuple use in transformed parallel region
• ** #5268 - Implement minmax support for Boolean type

Closed PRs

• #5281 - Make IR ir.Del free until legalized.
• #5280 - Added some hints for debugging entry points.
• #5278 - fix error messages in test_array_reductions
• #5276 - Add prange variant to has_no_side_effect.
• #5272 - Refactor parfor: split up ParforPass
• #5271 - Wip/rm dels 3
• #5270 - fix 'see also' in typeddict docs
• #5265 - Move preparfors pass

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

• Requests for 0.49
  • TBD