Minutes_2022_04_26

Numba Meeting: 2022-04-26

Attendees: Siu Kwan Lam, LI Da, Nick Riasanovsky, stuart, Shannon Quinn, Andre Masella, Benjamin Graham, Guilherme Leobas, brandon willard, Luk, Graham Markall,

NOTE: All communication is subject to the Numba Code of Conduct.

Please refer to this calendar for the next meeting date.

0. Feature Discussion

• llvmlite clang-format https://github.com/numba/llvmlite/pull/831

• mission statement discussion deferred to next week

• #8000 discussion

  @njit(locals={"s1": Series, "s2": int64})
  def foo():
      s1 = pd.Series([1,2,3,4], index=list("ABCD"))
      s2 = pd.Series([1,2,3,4], index=list("AAAD"))
      return s1, s2

  • if pd.Series always returns Series, locals can force casting to other types (e.g. int64) before storing into variables

• TODO: mutation of global reference to typedList segfault

  • https://github.com/numba/numba/pull/8000/files#diff-e41cdb72772b2d1985fcd9278c77c72c36c6e8d44aed8c86f6b9dc1a8148c70bR24-R25

• Emit object files from Numba AOT compiler https://github.com/numba/numba-extras/pull/4

• Discussion related to https://github.com/numba/numba/pull/6406

def foo_maker():
    a = 1
    @njit
    def foo():
        return a
    return foo
        
def foo_maker_2():
    a = 1
    exec(foo_source, {"a": a})
    foo = njit(<load_foo>)
    return foo

1. New Issues

• #7993 - optimization for frequently used code path
• #7995 - Numba 0.55.1 runtime error with numpy 1.21.6
• #7997 - Indexing a tuple/list of function raises an error
• #7998 - Allow registering and using extra jit options in Numba
• #8002 - CFuncs cannot use numba.carray when jitting is disabled
• #8005 - CUDA print tests do not work on Windows with recent drivers
• #8006 - Call to cuMemFree results in CUDA_ERROR_LAUNCH_TIMEOUT

Closed Issues

2. New PRs

• #7994 - Supporting multidimensional arrays in quick sort
• #7996 - Fix binding logic in @overload_glue.
• #7999 - Remove @overload_glue for NumPy allocators.
• **** #8000 - Add flag_class argument in jit decorator
• #8001 - Testhound/fp16 math functions
• #8003 - Add np.broadcast_shapes
• #8004 - CUDA fixes for Windows

Closed PRs

• #7992 - Pin llvmlite=0.39.0dev0 builds to build 63 on OSX

3. Next Release: Version 0.56.0/0.39.0, RC,