Minutes_2020_04_07

Numba Meeting: 2020-04-07

Attendees: Siu, Stuart, Todd, Guilherme, Aaron, Graham, Pearu, Ehsan, Val

0. Feature Discussion

• Whether to do 0.49RC2
  • Yes
  • April 9 Thurs
  • Target release on 16th.
• When to target 0.50 RC
  • burn down starts on 14th
  • cut RC on May 28th
  • final 4th June
• Graham on Czar duty this week
  • Stuart will cover Monday.
• Stuart: Doing flowchart of self-help triage system.
• Ehsan - ref leak thing?
• OMP: omp_set_nested

1. New Issues

• #5511 - for loop is not getting faster in some cases
• #5507 - Record mutation is ignored
  • record write opt-away
  • or known design bug from 2013
• #5503 - Code generation for np.einsum (including path optimization)
  • interesting
  • contain long code
• #5502 - ZeroDivisionError if an empty array is passed to np.mean()
  • draft patch https://github.com/numba/numba/pull/5512
  • also a problem w/ error-model
• **** #5501 - Post SSA: type unification error
  • Ehsan thinks accurate report on type-stability is better.
  • Also, better error message (currently hidden in the issue by NUMBA_DEVELOPER_MODE)
    • link to type unification page
• #5499 - Operations over a custom object aren't fused
  • related to the other overload operators & np.ufunc problem
  • parfor problem
• #5493 - Regression 0.49rc (Typing error)
  • fixed in 5495
• #5490 - support for with ... as ... expression
  • something went wrong in a PR
• #5486 - ModuleNotFoundError: No module named 'numba.config' in 0.49
  • shim PR in progress
• #5484 - Compatibility shims raise cryptic error
  • shim
• #5482 - Regression 0.49RC1 vs 0.48 (possibly SSA related)
  • SSA null init, fixed in PR #5495
• #5477 - literal_unroll fails when mixing iteration over untyped list
  • Fixed in #5477
• **** #5476 - Raising of an exception in inlined function fails on compilation
  • Stuart needs to talk to Ehsan
  • Fine to do.
• #5475 - numba loses its edge when dealing with large matrices
  • performance issue but can't reproduce
• #5474 - Treat typing.cast as a no-op?
  • feature request
• #5472 - literally with default values
  • check this against https://github.com/numba/numba/pull/5478
• #5471 - Default value is not literal
  • patch https://github.com/numba/numba/pull/5478
• #5470 - Regression 0.49RC1 vs 0.48 (tuple of functions)
  • patch https://github.com/numba/numba/pull/5473
  • related to first class function
• #5468 - typed.Dict.get return value has OptionalType when default value is provided
• #5466 - 0.49 RC: AttributeError: module 'numba' has no attribute 'numpy_support'
  • shim
  • patch 5497

Closed Issues

• #5508 - [0.49.0-rc1] (was #5413) numba.errors.RedefinedError: ... (step: analyzing bytecode) $const_0.5
• #5500 - heap implementation doesn't support typed list
• #5496 - Unable to use np.zeros
• #5487 - AttributeError: module 'numba' has no attribute 'numpy_support' in numpy 0.49
• #5483 - typed List() append changes types
• #5479 - Installation fails on Windows 7 x64 with Python 3.8.2
• #5469 - Capturing freevar tuples as tuples of consts
• #5467 - Change new np submodule name to the full numpy name

2. New PRs

• #5512 - Fix #5502
• #5510 - Creation of record array from list of tuples
• #5509 - allow the implementation of contexts which can type uncompiled dispatchers
  • needs checking w/ the author
• #5506 - Jitclass static methods
• #5505 - Better error message if __init__ returns value
• #5504 - Experiment type inference changes to fix 5501
• #5498 - Make the shim deprecation warnings work on python 3.6 too
• #5497 - Made numba.numpy_support available without an import
• **** #5495 - Fix missing null initializer for variable after phi strip
• #5492 - do not link unreferenced environments
  • need to ask for usecase
• #5491 - WIP: fix 'with ... as ... :' expressions
• #5489 - docs: fix rendering of nested bulleted list
• #5488 - Add missing numba.config shim
• #5485 - Show the offending module in "no direct replacement" error message
• #5481 - CUDA: Replace macros with typing and lowering implementations
• #5480 - Fix for #5477, literal_unroll KeyError searching for getitems
• #5478 - Fix #5471. Issue with omitted type not recognized as literal value.
• #5473 - Fix regression in 0.49RC1 for first-class functions

Closed PRs

• #5494 - Fix/5484

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