Skip to content
Valentin Haenel edited this page Apr 13, 2021 · 1 revision

Numba Meeting: 2021-04-13

Attendees: Todd, Graham, Leo, Jim, Luk, Kalyan, Stu, Siu, Val

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

0. Feature Discussion/admin

  • Should we have a official video channel (e.g. youtube) for numba tutorials and/or other educational materials.
    • MLIR keeps there meetup videos under google-drive
    • group ownership/management?
  • Internal error are being eaten by Numba
    • errors that shouldn't be suppressed are hidden by numba e.g. typo, invalid syntax
    • an overload version get accidentally skipped because of these errors
    • need changes to how code is raising exceptions to tell apart recoverable typing error vs critical error
    • may potentially break existing code that uses python builtin exceptions to indicate recoverable error
  • We should target 0.54 as after we stabilize all the in-flight target/hardware extension
    • 0.54 to support easier target extension
      • prove that new/existing target can use overload
    • fastmath flag needs rework
      • old way: LLVM module rewritten
      • new way should affect callees

1. New Issues

  • #6916 - Failing to build a jitclass with a deferred_type typed container as class member
  • #6908 - Tracker for target extension API potential issues
  • #6907 - Missing conversion from StringLiteral to unicode_type
  • #6906 - CUDA: Support __activemask
  • #6905 - Support error_model in cuda.jit
  • #6904 - CUDA: Produce a NumbaPerformanceWarning whenever a kernel is called on a host array
  • #6903 - Type inferring bug
    • Codegen could have unwrapped the optional type to see if it contains a valid value - if not then raise.
  • #6901 - Stencil neighborhood option doesn't support numpy.int64.
  • #6900 - divmod should return integers when passed unsigned integer and positive constant
  • #6897 - x in ... resolvable at compile time behaves inconsistently between cuda/njit
  • #6896 - Closures guaranteed to resolve are not handled well by inliner

Closed Issues

  • #6910 - Missing cudart_version_minor if CUDA RUNTIME can't load
  • #6909 - numba cannot resolve valid cuda driver
  • #6902 - sum(list())

2. New PRs

  • #6918 - Start section on commonly encounted unsupported parfors code.
  • #6917 - Add type conversion for StringLiteral to unicode_type and test.
  • #6915 - Update _Intrinsic for sphinx to capture the inner docstring
  • #6914 - Fix issue #6895 - walrus operator
  • #6913 - Add runtests option to gitdiff on the common ancestor
  • #6911 - CUDA: Add support for activemask(), lanemask_lt(), and nanosleep()
  • #6899 - Patch for support numba-dppy with device_context
  • #6898 - WIP: Work on overloading by hardware target.

Closed PRs

  • #6912 - Prevent use of varargs in closure calls.

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

Clone this wiki locally