Minutes_2021_04_06

Numba Meeting: 2021-04-06

Attendees:

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

0. Feature Discussion/admin

• 0.53.1 no reported (new) issues

• Implementing new features as per cpython and acknowledging protocols.

  • Example:
  • hash()
  • type.__hash__()
  • Existing example:
  • str()
  • Graham: a difficult situation
    • cuda array interface

  def foo(x):
      hash(x)
      
  @overload(hash)
  def ol_hash(x):
     def impl(x):
        return x.__hash__()
     return impl
     
  @overload_method(mynewtype, '__hash__')
  def ol_mynewtype_hash(thetype, ):
    def impl(thetype):
       return "this is a poor hash"
    return impl

  @overload(str)
  def ol_str(x):
     if query_type_has(x, "__str__"):
         def impl(x):
            return x.__str__()
     else:
         def impl(x):
            return "<default .....>"
     return impl

  • other point for consideration:
    • avoid spelling it twice/in-two-place
      • just having @overload_method should give all the information to query for a method in a type.
    • target specific?

• discussion evolved into purity in typing vs informative user error for a specific target

  • Data type is pure (the same across targets)
  • Function type is impure (depends on whether the target registered and implementation, and what's inside it)
  • In other words: typing is pure in the data model, but not in functions

• Typing context will have to care about the hardware because implementation will potentially cross a function boundary

  • However a lot of this will be generic and be able to use generic implementations anyway
  • e.g. __str__(), __hash__(), __len__(), etc., the same on all targets

• "Initial support for Numpy subclasses' https://github.com/numba/numba/pull/6148

1. New Issues

• #6895 - := operator cannot be jitted by at least the cuda module
• #6894 - closure inlining not dealing with star-args correctly
• #6893 - Catch incompatible cudart arch
• #6892 - Numba and tbb version 2021 packages incomatibility on Linux
• #6891 - Cuda functions treat *args as a normal parameter
• #6887 - Allow setting across slices of slices
• #6884 - CUDA matrix multiplication example is probably wrong
• #6882 - LoweringError using njit(parallel=True) when using python 3.8 (but not when using 3.7)

Closed Issues

• #6875 - Issue with __synchronize() identifier undefined.

2. New PRs

• #6890 - fixes #6884
• #6889 - Address guvectorize too slow for cuda target
• #6888 - Get overload to consider compiler flags in cache lookup
• #6886 - CUDA: Fix parallel testing for all testsuite submodules
• #6885 - CUDA: Explicitly specify objmode + looplifting for jit functions in cuda.random
• #6883 - Add support of np.swapaxes #4074
• #6881 - support the 'reversed' function for ranges, lists, and arrays
• #6880 - Add attribute lower_extension to CPUContext
• #6879 - CUDA: NumPy and string dtypes for local and shared arrays
• #6878 - CUDA: Support passing tuples to ufuncs
• #6877 - Add doc for recent target extension features
• #6874 - Clarify conversions to numeric types

Closed PRs

• #6876 - Add trailing slashes to dir paths in CODEOWNERS

3. Next Release: Version 0.54.0/0.37.0, RC=May 2021