Minutes_2020_06_23

Numba Meeting: 2020-06-23

Attendees: Todd, Val, Stuart, Pearu, Aaron, Siu, Guilherme, Hameer Apologies: Graham

0. Feature Discussion

• 0.50.1 Wed

  • BUG: get_terminal_size fail in some terminal
  • BUG: deprecation notice bump
  • BUG: inliner issue w/ literally
  • BUG: error messages problem w/ CUDA target

• discourse feedback?

  • issue template updated to redirect to discourse
  • add sub-category: https://numba.discourse.group/t/testing-with-a-suggestion-on-site-organization/28

• Numba release candidate and llvmlite

  • Pearu reported pip packaging conflict but now it works
  • probably caught between wheel uploading before it finishes

• Guilherme questions:

  • @guvectorize compile new versions
  • sound like generalize dufunc to work like guvectorize
  • Siu suggest short term workaround maybe to wrap guvectorize function to generate new gufunc for each new signature.
  • Siu thiks this is good work to attempt but not easy.

1. New Issues

• #5902 - cuda.jit does NOT preserve the original __doc__ and __module__ of the decorated function
• #5901 - cuda: inconsistency between data type specifications in device_array and local.array
• #5899 - Add support for programming Tensor Cores in CUDA kernel
• #5898 - isinstance fails with complex numba types
• #5897 - Reflected-list and typed List produce slow concatenate for list of equal sized arrays
• #5896 - 'NoneType' object has no attribute 'args'
• #5895 - Support for numpy.dot in higher dimensions
• #5894 - Support for np.ix_
• #5891 - Weird warnings when using @jit on version 0.50
• #5890 - Seemingly random segfault on macOS if function is in larger library
• #5887 - Regression with literally and overloads with inline='always'
  • has a patch
• #5885 - Add support for overloading __call__ on custom types with extending.overload
• #5883 - Reminder to remove the numba.jitclass shim
• #5881 - Cannot run on simple for loop
• #5880 - Bug: Argwhere returns a storage
• ** #5878 - ConcurrentBag-like data structure for parallel appends
  • Todd:
    • thinks it's not far off
    • does not involve array_analysis
    • gufunc limitation will be a bigger issue
  • Siu:
    • bigger feature to tackle on gufunc replacement/updates
• ** #5875 - libdispatch-based (GCD) workqueue
• #5874 - Support literal string in structured dtype
• #5873 - Implement type key-word-argument for numba.typed.List constructor
• #5872 - cannot write to numpy array in namedtuple with parallel=True
• #5871 - Bug on pyapi.call when args is None
• #5870 - np.linalg.pinv Unintended mutation of input
  • has patch

Closed Issues

• #5888 - how to specify the number of threads when using prange

2. New PRs

• #5900 - Enable CUDA Unified Memory
• #5893 - Allocate 1D iteration space one at a time for more even distribution.
• #5892 - Better handling of tuple arguments whose fields are modified.
• #5889 - Fixes literally not forcing re-dispatch for inline='always'
• #5886 - Add support for overloading the call operator
• #5884 - Update the deprecation notices for 0.50.1
• #5879 - Fix erroneous input mutation in linalg routines
• #5877 - Jitclass builtin methods
• #5876 - Improve undefined variable error message

Closed PRs

• #5882 - Type check function in jit decorator

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

• Opening up the numba meeting