Minutes_2020_01_21

Numba Meeting: 2020-01-21

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

0. Feature Discussion

• 0.48.0 release status

  • Still need to tag 0.49.0dev0
  • build on Friday

• __cuda_array_interface__ https://github.com/numba/numba/issues/4933#issuecomment-572934894 - choices (more than 1 is possible) - a) safe default with implicit sync - b) unsafe default and let user sync - c) new interface to provide a) or b)

  • Graham noted that if there's e.g. aynchronous memory management on streams at some point this will have to be addressed.
  • Take default current behaviour to unblock this?
  • Question:
    • What's the actual use case
      • TF using Numba in its graph
      • Numba thing, then TF, then Numba thing

• Quite a few users last week or so using Inteval example, it has problems, is missing things, and makes assumptions/does invalid stuff. Need to update lower level extension API.

  • https://gist.github.com/luk-f-a/6d0eef6534a8aae27a6650093321d833

• First class functions

  • try get this in this in this week before code movement changes
  • Open Qs:
    • Function addresses baked in? Global function, maybe!? Need to check. Resolve this on PR.
    • Usability issues, report and do follow up PRs.

• Thread masking

  • ready for review

• Chainsawing

  • remove py2.7, py3.5

• File moving plan

• Actions from last meeting:

  • Ask Siu about Function type PR going into 0.48: Done
  • Stu look at radare2 for asm CFG: Done

• TODO:

  • Share plan of chainsawing
  • ✔️ Siu/Stu ask Q on #4933
  • TODOs for others orthogonal to chainsaw
  • Siu #4967 replace inttoptr with annotation for caching

1. New issues

• #5069 - Improve error message for accidental use of a non-void function spec in @guvectorize.
• #5070 - is there any Thread Safe Container ?
• #5071 - Segfault importing numba
• **** #5072 - Accept prange driven loops in the loop body versioning transform.
  • No not until tuple problem is fixed.
• #5073 - Slices of dynamic shared arrays all alias
• ** #5074 - Readability issues in the docs
  • old theme in use at the minute
  • RTD theme?
  • Change CSS colours for notes/warnings
• #5075 - np concatenate slower in numba
• #5079 - Further problems with literally in overload
• #5081 - typed.Dict values() returns wrong results when dict key_type is np.int64
• #5083 - Lowering error using the IntervalType example from the documentation
• #5084 - Support for Stack-Allocated Arrays
• #5086 - np.nan is np.nan behaves differently when jitted
• **** #5087 - Function working with 0.46 doesn't work with 0.47 (runs indefinitely)
  • really hard to figure out problem with certain bytecode
  • Port over to 0.48.0
• #5088 - Support the sum builtin
• #5093 - warning: ignoring debug info with an invalid version (0) in <string> with strange behaviour
• #5094 - object mode fails sometimes
• #5095 - In jitclass, will not compile when a member variable is used in numpy.zeros()
• **** #5096 - Compilation time of a njit function getting too long when one parameter is a large list of jitclass
  • ref count
  • TODO ticket: timers on dispatch + boxing would be useful
• #5097 - "pass" combined with isnan fail in njit mode in 0.47
• **** #5098 - _lower_parfor_parallel function can mess up sequential_parfor_lowering flag disabling parallelization for the rest of the execution
  • Todd to take a look

• **** #5100 - Bug with addition assignment operator in an array when using njit(parallel=True)

• #5101 - Add @lower_constant example to IntervalType doc
• #5103 - Option to run Numba tests with pdb
• #5105 - Unification failure with NaT
• #5107 - Imprecise result when raising complex128 to power of complex128
• #5108 - modifying numpy in the paralel loop
• #5109 - Numba nested function won't execute (version 0.47)
• #5110 - feature request, more supported forms of np.round
• #5112 - Invalid use of Function(<built-in function setitem>)
• #5113 - Fix error handling in Interval example

Already Closed

• #5082, #5080, #5077, #5076, #5070

2. New Open PRs

• #5085 - Support other itemsizes with view
• #5092 - Fix 5087
• #5099 - Fix #5073: Slices of dynamic shared memory all alias
• **** #5102 - IR node aliasing check
  • WIP
• #5104 - Add a list of libraries that support cuda_array_interface
• #5106 - Add SciPy 2019 talks to docs
• #5111 - [WIP] lowering of constant jitclass

Already merged/closed

• #5058 - Permit mixed int types in wrap_index

4. Next Release: Version 0.49.0, RC=March ???

• Requests for 0.49
  • TBD