Minutes_2020_04_14

Numba Meeting: 2020-04-14

Attendees: Ehsan, Graham, Pearu, Stuart, Todd, Val

0. Feature Discussion

• RC2 and official release status
  • entrypoint test failing on windows on the farm
    • cannot reproducer on public CI
    • probably problems in conda
  • power tests failing
  • need to check integration test against master
• Gitter now has threaded conversations
  • editing code blocks don't work
• https://github.com/numba/numba/pull/5512 <-- what should we do here
• Pause on overload refactoring because declarative type
  • this lists all @lower_builtin http://numba.pydata.org/numba-doc/latest/developer/autogen_lower_listing.html
• declarative typing WIP:
  • issue for this task: https://github.com/numba/numba/issues/5552
  • what error messages should look like
  • what can we do to write this and also not break existing things
  • how can the scope be limited to make improvement achievable
  • add debug tools
  • Pearu proposes that the errmesg can suggests how a missing signature could be implemented.
    • different group of users/devs have different needs from errmesgs.
    • different verbosity-level or flags would be needed
• https://github.com/numba/numba/issues/5424
  • probably got fixed by SSA

1. New Issues

• #5550 - No function '__pyx_fuse_0pdtr' found in __pyx_capi__ of 'scipy.special.cython_special'
  • has a fix
  • scipy signature drifted
• #5545 - Implement math.modf for CUDA target
• #5541 - Fixed-size stack-allocated list
  • Could use tuple to emulate this? (TODO)
• **** #5540 - named parameters not recognized with first class functions
• #5539 - Invalid np.dot() output for vector/matrix multiply of empty arrays
• #5537 - platform.linux_distribution is deprecated
  • patched in #5537
• #5536 - Implement str correctly
• #5535 - using numba with "from collections import Counter"
• #5533 - data type always meets bugs
  • comparison operator on dtype
• **** #5531 - Accessing generators from c extension
  • unsupported use
  • would need a lot of work to extend generator
• #5530 - CUDA datetimes are only tested with datetime[D], documentation not explicit
• #5525 - Add figure for total memory to numba -s output.
• #5524 - How to create a typed dict in Numba dict[str] = list
• #5523 - np.ravel does not preserve mutability for class attributes with @jitclass
• #5522 - Tuple of inner functions typing regression in 0.49rc1 due to first class functions feature
• #5520 - OMP omp_set_nested deprecated warning in 0.49rc1
• **** #5515 - Invalid copy of Array analysis object for IR extension handlers (regression in 0.49rc1)
  • It's unclear how information is returned/mutated by ArrayAnalysis. Can we improve this?
  • Ehsan suggested that we can merge instances of ArrayAnalysis to avoid the current indirect mutation of attributes.
  • Todd reminded us that ArrayAnalysis is used a lot and may affect compilation time greatly for parfor-enabled code.
  • Ehsan reminded us that compilation time profiling is made difficult because of how numba compiles recursively (i.e. typeinfer trigger new compilation)
  • Stuart suggests we can measure at the pass-machinary
• #5513 - Recursive inline=always functions never stop compiling

Closed Issues

• #5547 - numba.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend) Type of variable 'argmax' cannot be determined
• #5528 - Numba 0.49.0rc1 does not provide aliases for some use cases
• #5521 - Invalid use of Function(<built-in function setitem>)
• #5518 - problem typing a global tuple of specific integer types as literal
• #5516 - release 0.49.0rc2 checklist

2. New PRs

• #5549 - Check type getitem
• #5548 - [WIP] Fix #5537 Removed reference to platform.linux_distribution
• #5546 - DOC: Add documentation about cost model to inlining notes.
• #5544 - Add support for np.union1d
• #5543 - [WIP] np.trim_zeros
• #5542 - Fixes RC2
• #5526 - Impl. np.asarray(literal)
• #5519 - CUDA: Silence the test suite - Fix #4809, remove autojit, delete prints
• #5517 - Fix numba.typed.List extend for singleton and empty iterable
• **** #5514 - Code generation for np.einsum
  • style is unlike the rest of numba
  • source generation style is very complicated
  • check how numpy implemented np.einsum
  • no decision on how to consume this

Closed PRs

• #5538 - Update CHANGE_LOG for 0.49.0rc2
• #5534 - Commit to bump test_entrypoints to get a CI run
• #5532 - Make numba.<mod> available without an import
• #5529 - PR #5473 continued
• #5527 - Attempt to fix #5518

3. Next Release: Version 0.50.0, RC=???

• 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