Minutes_2019_10_15

Numba Meeting: 2019-10-15

Attendees: Ehsan, Pearu, Val, Stan, Siu, Todd, Stuart, Aaron

0. Feature Discussion

• Numba 0.46.1

  • Python 3.8 was released on Monday
  • Critical bugs
    • inliner bug @stuart

• Copied from last 2 weeks

  • Requests for 0.47 (last release for the year)
    • jitclass performance issues
    • llvm 9 trial
    • CTK libcudadevrt.a
    • CI needs to take 50% of current time
      • Val & Stu already looking at this
      • also checking Azure CI config to avoid wasting compute time
    • Caching:
      • transitive dependency
      • other issues: i.e. function as argument, with objmode
      • distributing cache
    • Immutable list and deprecating reflected list
    • Switch to pytest (see above)
    • Using Numba to generate LLVM/NVVM IR for different targets https://github.com/numba/numba/issues/4546
      • @overload for gpu

1. New issues

• #4702 - Problem installing numba 0.46 and numpy 1.17 in conda

  • we still build with numpy pinned in farm
  • official builds are not affected
  • new CI farm is not affected

• #4701 - Numerical differences when using numpy.linalg.norm

  • numpy bug?
  • add FAQ entry on floating point precision

• **** #4698 - Serializing typed Lists and Dicts?

  • pickle-able Lists and Dicts make sense
  • Stan thinks:
    • focus on portability not performance
    • make pickle work
  • Val thinks:
    • require reasonable performance
  • Need more thinking.

• #4697 - nested dictionaries failed

• **** #4696 - Inconsistent behavior of stencil decorator

  • parfors
  • Todd will investigate

• #4694 - Default None argument treatment is different in type-inference and in lowering

• IR inliner/rewrite-passes related

  • Issues:
    • **** #4693 - raising Exceptions in inlined functions fails on compile
      • Need to introduce pass grouping?
    • **** #4692 - generated_jit doesn't get passed correct arguments with inline='always' argument
      • Should generated_jit be deprecated in favor of overload?
        • or just merge the impl
    • **** #4691 - tuple slicing fails in inlined functions
  • Action
    • Issues are not critical
    • Do better fix as suggested by Ehsan
    • better separation of analysis vs transformation

• #4690 - numba 0.46, useless(?) code in parfor init block

  • waiting for reproducer

• **** #4689 - The fastmath SVML option has been broken since somewhere between 0.43.1 and 0.44

  • initializing the LLVM engine before SVML option is set will freeze the no-SVML setting for subsequent compilations
  • not detected by unit tests because compile_isolated creates a new LLVM engine every time

• #4688 - Add support for @classmethod in jitclass

  • put on the jitclass wishlist

• #4687 - inline_closure_call may ignore glbls parameter

  • replied

• #4684 - Missing arg in TypeError

• #4683 - Help for VMProf with Numba

  • need more information
  • maybe point them to numba/stacktrace?

• #4681 - Error in py_call_impl(callable, dots$args, dots$keywords)

  • need more info

• #4679 - Numba 0.46.0 final checklist

  • closed

• **** #4676 - CUDA decorator/compiler argument to set literal precision

  • need to fix the integer case in 0.47
    • see NBEP
  • what about float?
    • literal only?
    • __truediv__(int, int) or math.log(int)->float??
    • what does C do?

• #4674 - Same code works on macOS Python 3.7.2 but not on Ubuntu 16.04 Python 3.5.2

  • Siu looking into to this.
  • Linux py3.5 only bug so far.
  • problem in IR mutation in parfor transformation

• #4671 - can't pickle DUFunc objects

  • see PR

• #4670 - I have a question about half precision float support in numba cuda

  • only just started

• #4668 - Round segfaults in cuda jitted function if passed ndigits optional parameter

  • need to catch this in typing

Already Closed

• #4700 - Numba makes Large Numbers Negative
• #4699 - Request for f-string support and numpy.isscalar.
• #4686 - Cannot parallelize a loop
• #4678 - @jit decorator with signature including List(List(int64)) not working

2. New Open PRs

• #4703 - Fix numba.jit parameter name signature_or_function
• **** #4695 - Refactor overload* and support jit_options and inline
• #4677 - Add support for np.setxor1d
• #4673 - Extend test timeout and add identifier cmdline
• #4672 - Fix pickling of dufunc
• #4669 - Add link to ParallelAccelerator paper.

Already merged/closed

• #4685 - Apply #4682 to 0.46 release branch
• #4682 - Update changelog for 0.46 final release
• #4680 - Apply #4675 to on 0.46 release branch
• #4675 - Bump cuda array interface to version 2

4. Next Release: Version 0.46.1, RC=

• CPython 3.8