Minutes_2020_08_18

Numba Meeting: 2020-08-18

Attendees: Siu, Ehsan, Graham, Guilherme, Hameer, Stuart, Todd, Val

0. Feature Discussion

• 0.51.0 post release status

  • wheel issues in relation to DSO bundling
    • TODO ask PyPA
  • Performance regression
    • No reproducer yet for 0.50.0->0.51.0
    • There are report from large code project.
    • Users are trying to reduce and pin point the problem.
    • Reasons:
      • SSA:
        • started in 0.49
        • compute intensive algorithm
        • overly eager computation of dominance
        • too much re-computation of CFG info
      • namedtuple:
        • patched in 0.48.0, 0.49.0
        • patched need to fix dispatch (#6145)
  • 0.51.1
    • objmode cache segfault (#6130)

• Compilation Speed

  • Ehsan asked if we should consider C/C++/Cython version of IR and some of the expensive algorithms
    • Stuart concerns about maintenance burden of C++ code and we should fix excessive SSA passes before actually moving to C++.

• Ehsan's team

  • has more cache crash issue to report
  • enhancement to with objmode() to use type references instead strings

• 0.52 Working On

  • Improving compilation speed
    • fixing smaller obvious problems
    • refct pruning in LLVM C++ pass
  • Graham align CUDA dispatcher to look more like CPU dispatcher
  • Guilherme: dynamic ufunc pr #5938 (TODO needs review)
  • Todd: NumPy subclasses

• Hameer: typedlist performance regression comparing to reflectedlist

  • Stuart:
    • would be partly fixed by new refct pruning
    • other inefficiency in codegen
  • Val suggested potential workaround by fallback to reflectedlist

• longer RC period by default

  • 8 weeks + 2 (uncertainty) = 4 (dev) + 2 (burndown) + 2 (rc) + 2 (uncertainty)

• themes Q3 + Q4

  • compilation speed improvements
  • diagnostics for this

• discourse interesting things?

  • https://numba.discourse.group/t/numba-access-of-numpy-ufunc-inner-loop-functions/
    • numpy changes ufunc C-API
    • Does numba use ufunc innerloop? we are not sure
    • IMPORTANT change in NumPy
    • Is this change additive? Would mean Numba can migrate to it as resources allow.

• TOPICS WE COULDN'T GET TO

  • opt-in features

1. New Issues

• **** #6136 - Compilation time regressions 0.47-0.51.
  • We'll need to revive the perf suite
• #6135 - Unicode characters as dict keys don't work reliably due to character width ambiguity
  • interesting string slicing details
• #6134 - mssa.fit()出错
  • lack info; duplicate
  • original issue may have been caused by numba bugs
  • may have been fixed on latest
• #6133 - Compile times are nontrivally higher in 0.50 vs 0.46
  • more arguments for caching cuda kernels
• #6131 - cupy + numba cuda error: [304] Call to cuInit results in CUDA_ERROR_OPERATING_SYSTEM
• #6130 - Cache does not work when passing numpy arrays to objmode inside overload
• #6125 - IndexError in _version_info with build from master
  • patched in #6126
• #6118 - support np.expand_dims second tuple argument
• #6114 - Setting an item in a structured type in a device array behaves differently than raw arrays.
  • __setitem__ bug on device record array
• #6113 - Fine-tuning of calling external function pointer
• #6108 - TBB can't be loaded
  • TODO: ask PyPA

Closed Issues

• #6138 - PicklingError while returning typed list of jitclass objects from njit function
• #6132 - SIGSEGV with jitclass + typed.Dict
• #6121 - CUDA: Recent commit causes infinite loop (or a lot of recompiling)
• #6119 - Release 0.51.0 checklist
• #6106 - Building/Using Numba with PTDS enabled
• #6105 - What's the input of Numba

2. New PRs

• #6140 - Replace calls to utils.iteritems(d) with d.items()
• #6139 - Use direct syntax instead of the add_metaclass decorator
• #6137 - Remove references to Python 2's long
• #6129 - Extend Cudasim to support most of the memory functionality.
• #6128 - CUDA Docs: Restore Dispatcher.forall() docs
• #6127 - Fix race in reduction kernels on Volta, require CUDA 9, add syncwarp with default mask
• #6126 - fix version_info if version can not be determined
• #6122 - Make strides is None C-contiguous
• #6115 - Add libtbb license info for wheel packages
  • TODO Rebuild old numba wheels (0.50.1?) to strip tbb/gomp
• #6112 - CUDA: Make block, thread, and warp indices unsigned.
  • Note: consider index_type as an integer without cast to float semantic
• #6110 - Update CONTRIBUTING.md

Closed PRs

• #6124 - Update CHANGE_LOG for #6123
• #6123 - Fix #6121
• #6120 - bump llvmlite to 0.34
• #6117 - Update CHANGE_LOG for 0.51.0
• #6116 - Fix #6102. Problem with non-unique label.
• #6111 - Decouple LiteralList and LiteralStrKeyDict from tuple
• #6109 - Fixes an issue reported in #6094
• #6107 - Fix and test #6095

3. Next Release: Version 0.52.0, RC=?, Final ?

• Requests for 0.52
  • Fast(er) typed.List/typed.Dict? Doesn't have to be in 0.52, next 3-5 months is fine.

4. Upcoming tasks