Minutes_2020_04_28

Numba Meeting: 2020-04-28

Attendees: Perau, Guilherme, Aaron, Graham, Todd, Val, Siu, Stuart

0. Feature Discussion

• numba 0.49.1 and llvmlite 0.32.1

  • Potentially cut RC1 for Thurs

• Exceptions:

  • Want to "fix" open ended...
    • At present exceptions have to be "constant".
    • Want to make exception string?/args to be dynamic.
    • Could Numba pickle a function that raises the exception?
      • Hard parts:
        • C code

• Replace inline_closurecall (maybe)?

• (from prev discussion) Q&A forum that's not gitter or mailing list

  • gitter:
    • high traffic, hard to search, threads in beta
    • we should report issues
  • StackOverflow:
    • low signal to noise ratio
    • may have improved; "numba" tag working
    • Very focused only on specific questions with specific answers - could be a problem?
  • mailing list:
    • low traffic, old
  • Discourse (updated)
    • may need investigation
    • try out within this group
    • https://discuss.python.org/
  • SO (updated)
    • try observing SO "numba" tag for a little
    • SO QA style doesn't fit our need
    • we need more discussion and questions may have a clear answer

1. New Issues

• **** #5625 - function with parallel forloops + njit raises error (inconsistently)
  • parfor does not work with Fortran arrays
  • try to catch and make better error message
• #5623 - SSA pass causes invalid compilation (array write elided).
• #5622 - Numba 0.49.1RC1 checklist
• **** #5621 - Nested prange regression in 0.49
  • for TBB experts: Any quick win here?
• #5618 - Jit produces wrong code with version 0.49.0
• #5615 - First class function type breaks literal_unroll
• #5613 - np.floor_divide does not match numpy for divide by zero with floats
  • numpy bug
• ** #5611 - Control flow graph algorithm error when using many ternary operations
  • interesting autogen usecase
  • Stuart's Q: compiling the compiler
    • could have JIT overhead
    • caching or AOT may be needed
    • 0.49.1 - Don't compute CFG for each
• ** #5608 - Typecode drift
• #5605 - Task: move examples to numba/numba-examples
• #5603 - Numba List fails with differently named functions (even if signatures match)
• #5601 - Failure in step: convert to parfors
• #5600 - ImportError: Numba could not be imported
  • How do we stop having this problem?
    • add test using lief
    • check compiler version
• #5599 - Lowering error with Numba 0.49.0
• #5597 - Invalid IR generated in Parfors
• #5596 - cast via NumbaClass (i.e np.intp) incorrect for array input
• #5595 - CUDA: Provide memset (or memset-like) functionality as part of the public interface
  • a contributor is working on this
• #5594 - Single float value in numba Dict

Closed Issues

• #5633 - Parallel example function error
• #5632 - ImportError: Numba could not be imported.
• #5631 - ImportError: Numba could not be imported.
• #5620 - Strange results using shared memory matmul example on website
• #5612 - Numba returns error when using numpy inverse function
• #5610 - ImportError: Numba could not be imported
• #5604 - Error in py_module_import(module, convert = convert) : ImportError: Numba could not be imported.
• #5598 - KeyError 'b'

2. New PRs

• #5630 - Fix #5570. Incorrect race variable detection due to SSA naming.
• #5629 - Fixes naming error in array_exprs
• #5628 - Fix using type to cast and make np.array() work on array
• #5627 - Fixes #5623, SSA local def scan based on invalid equality assumption.
• #5626 - Unify undefined first class function types [WIP]
• #5624 - Fix erroneous rewrite of predicate to bit const on prune.
• #5619 - Add support for multi-output ufuncs
• #5617 - Remove /examples from the Numba repo.
• #5616 - Simplify the ufunc kernel registration
• #5614 - Fix invalid type in resolve for comparison expr in parfors.
• **** #5609 - Jitclass Infer Spec from Type Annotations
  • View as a proposal to support Type Annotations in general
  • May need a lot of discussion
• ** #5607 - Add StackArray example
• #5606 - Support None as an out argument in ufuncs
• #5602 - Fix segfault caused by pop from numba.typed.List

Closed PRs

3. Next Release: Version 0.50.0, RC=May 28th

• Requests for 0.50

• high risk stuff for 0.50.

  • Declarative typing (Siu and Stu)
    • declaring accepted types in things like @overload
    • so we can have better errmsg
    • allow overloading "contextual information"; i.e. targets (CPU, GPU), fastmath flag, exception handling
    • make sure we don't break existing @overload/@lower_builtin use cases
  • Remove macro expansion (Graham)
    • replace "macros" with standard typing/lowering; i.e. @register / @lower (@overload not yet working with CUDA)

• llvm 9