Minutes_2024_03_05

Numba Meeting: 2024-03-05

Attendees: FPOC (last week): FPOC (incoming):

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

Please refer to this calendar for the next meeting date.

0. Discussion

• Numba user survey update.

  • 91 responses so far 🔥
  • https://numba.discourse.group/t/numba-user-survey-2024/2411
  • please continue to share 🙏

• Numba 0.59.1 schedule.

  • Aim for patches in this week, test build over the weekend and release next week.

• Review NumPy 2.0 community communication note.

• How to stage NumPy 2.0 work and impact on maintenance.

• Draft of February's NAN is ready for comments from maintainers.

• Possible "solutions" to the question last week about how to specialise compilation based on a configuration object, and still preserve literals.

  from numba import jit, types
  from numba.extending import overload
  import functools
  
  
  # Some sort of configuration object
  class Config():
     def __init__(self, a, b):
         self._a = a
         self._b = b
  
     @property
     def a(self):
         return self._a
  
     @property
     def b(self):
         return self._b
  
  
  # Perhaps use a cache so that the identical Config instances return the same
  # jit function? This will prevent recompilation of the entry point for two
  # identical config instances as the jit function passed as the argument will be
  # the same.
  @functools.cache
  def obj2strkeydict(obj, config_name):
  
     # unpack object to freevars and close over them
     tmp_a = obj.a
     tmp_b = obj.b
     assert isinstance(config_name, str)
     tmp_force_heterogeneous = config_name
  
     @jit
     def configurator():
         d = {'a': tmp_a,
              'b': tmp_b,
              'config_name': tmp_force_heterogeneous}
         return d
  
     # return a configuration function that returns a string-key-dict
     # representation of the configuration object.
     return configurator
  
  
  # Define some "work"
  def work(pred):
     # ... elided, put python implementation here
     pass
  
  
  @overload(work)
  def ol_work(pred):
     assert isinstance(pred, types.Literal)
     print('Calling work with type', pred)
     return lambda pred: pred
  
  
  @jit
  def physics(cfig_func):
     # This is the main entry point to the application, it takes a configuration
     # function as an argument. It will specialise on each configuration
     # function.
  
     # call the function, config is a string-key-dict
     config = cfig_func()
     # unpack config, these types will be preserved as literals.
     a = config['a']
     b = config['b']
     # call some work to check the types.
     return work(b) + a
  
  
  # demo
  def demo():
  
     # Create two different Python based configuration objects with literal
     # entries.
     configuration1 = Config(10, True)
     configuration2 = Config(12.3, False)
  
     # Create corresponding converted configuration objects...
     jit_config1 = obj2strkeydict(configuration1, 'config1')
     jit_config2 = obj2strkeydict(configuration2, 'config2')
     # create "another" configuration1 instance, memoization prevents
     # duplication.
     jit_config1_again = obj2strkeydict(configuration1, 'config1')
  
     # Call the `physics` application, it will specialize on config.
     physics(jit_config1)
     physics(jit_config2)
     # should not trigger a 3rd compilation, config is memoized.
     physics(jit_config1_again)
     physics.inspect_types()
  
  
  if __name__ == "__main__":
     demo()

• LLVM upgrade:
  • discourse thread: https://numba.discourse.group/t/llvmlite-upgrade-cadence/2436
  • For 15: Passes not supported on legacy pass manager:
    • ArgumentPromotionPass: https://reviews.llvm.org/D128536
    • LoopUnswitchPass: https://reviews.llvm.org/D124376
  • Effort to migrate to new pass manager estimates?
    • ???
  • Other tasks:
    • llvmdev recipe update
    • https://github.com/numba/llvmlite/tree/main/conda-recipes/llvmdev
  • Required patch when building LLVM for testing:
    • https://github.com/numba/llvmlite/blob/main/conda-recipes/llvm14-clear-gotoffsetmap.patch
  • Beyond LLVM 15:
    • Opaque pointer support
    • Other changes?

New "Ready for Review" PRs

1. New Issues

• numba#9471 - CUDA crashes when passed complex record array
• numba#9472 - [ANN] Numba User Survey 2024
• numba#9473 - want int64 rather than Optional(int64) when dict.get(key, default_int64) and dict value type is int64
• numba#9475 - Infer Numba Types from Python Type Hints
• numba#9476 - [feature request] assign a tuple to one row of recarray
• llvmlite#1034 - Add numba and llvmlite musl wheels for alpine support

Closed Issues

• numba#9477 - np.searchsorted no longer work for datetime64[ns] array type

2. New PRs

• numba#9470 - Add ability to link CUDA functions with in-memory PTX.
• numba#9474 - (Do Merge) NumPy 2: Removed dead code conflicting with binary builds against NumPy 2

Closed PRs

(last numba: 9477; llvmlite 1034)

3. Short-term Roadmap

2024-gantt: TBD 2023-gantt: https://github.com/numba/numba/issues/8971