Skip to content

Minutes_2020_06_02

Jump to bottom Edit New page
Siu Kwan Lam edited this page Jun 2, 2020 · 1 revision

Numba Meeting: 2020-06-02

Attendees: Aaron, Graham, Guilherme, Hameer, Pearu, Stuart, Todd, Siu

0. Feature Discussion

  • Openmeeting
    • CoC:
    • scheduling
      • June 9th (Tue)
    • Agenda ?
      • Intro's
      • Mailing list and gitter - specific Q's/topics etc.
      • "Normal" maintenance/triage/issues for a few minutes
  • RC1 release weds
    • LLVM 9 on master
      • CUDA broken
        • should be fixed by 3 PRs by Graham
    • Need to patch farm smoke_test_all to include CUDA/ROC
    • NumPy 1.18
    • Performance regression patch: https://github.com/numba/numba/pull/5795
  • Finer-grained labels (all open issues are at least labelled)
    • Stuart to start making a list
  • Declaring "support" for something e.g. NumPy version
    • Consider 1.18 bugs not hugely important vs. general support.
    • NumPy 1.19 may be ABI incompatible.
    • Some discussion about wanting to be able to express "we tested this" and "we didn't test this"
      • Conclusion:
        • No upper bound should be applied, matches up to other places
        • NumPy 1.18 is ok, just couple of bugs, don't feel need to restrict
        • Numba 0.50 out soon which will support NumPy 1.18 "officially"
        • can release a 0.49.x as a "upper bound" release if needs be

1. New Issues

  • #5792 - Not work jit_options={'error_model': 'numpy'}
    • needs more thought
  • #5790 - Support for EMM context managers
  • #5788 - under @njit, the result of np.max/np.argmax is actually np.nanmax/np.nanargmax
    • to be fixed in np1.18
  • #5786 - np.positive is not implemented
  • #5784 - missing lowering for static_setitem with slice on char array
  • #5780 - Setting a datetime value on a Numpy structured array throws an error
  • **** #5778 - prange reverts to serial when thread count doesn't divide iterations
    • see NUMBA_DEBUG_ARRAY_OPT for thread scheduling & check that scheduling is reasonable
  • #5777 - Printing a Numpy structured array containing a string column throws an error
    • null-terminators
  • #5775 - Additional work needed on the new inliner
  • #5774 - "access violation" problem
  • #5771 - jitclass test failure in longlabel_not_truncated: LLVM use of undefined value
    • due to using non-patched LLVM
  • #5770 - List populated during multiple loops does not convert to array correctly.
  • #5769 - Compiler explorer for Numba
  • #5767 - Fix up the removed refcount pruner for arrays, make it more strict but still work!
  • ** #5766 - PyPy support
  • #5762 - Adding get_memory_manager
  • #5760 - literal_unroll: allow setitem in loop; also allow zip, enumerate, et al. w/ literal_unroll?
  • #5755 - @vectorize-decorated functions are not found by sphinx
  • #5754 - Using shared memory in Numba with Cupy functions
  • #5753 - Declaring Numpy structured array dtypes throws TypeError
  • ** #5752 - NumPy 1.18 Support
  • #5750 - meta discussion: jitclass inferred constructor
  • #5749 - Using Exceptions to trigger continue in loops fails
  • #5748 - Support getitem for type types.
  • #5747 - Check if variable is None inside njitted function is neither True nor False with parallel=True

Closed Issues

  • #5787 - Segfault when accessing inferred typed.Dict object poped from typed.List.
  • #5776 - Error at driver init: [100] Call to cuInit results in CUDA_ERROR_NO_DEVICE:
  • #5772 - AttributeError: module 'numba' has no attribute 'targets'
  • #5756 - Indexing a structured Numpy array using a variable throws an error

2. New PRs

  • #5791 - Fix up no cgroups problem
  • #5789 - [WIP] support for np.putmask
  • #5785 - [WIP] Implement np.tile
  • #5783 - Make np.divide and np.remainder code more similar
  • ** #5782 - [WIP] Host docs on Read the Docs
  • #5781 - Moving record array examples to a test module
  • #5779 - Return zero array from np.dot when the arguments are empty.
  • #5773 - fix exception during locale detection
  • #5765 - [DON'T MERGE] MWE for sphinx not finding vectorized funcs
  • #5763 - Support minlength argument to np.bincount
  • #5761 - [WIP] Add implmentation for int types for isnan and isinf for CUDA
  • #5758 - Support NumPy 1.18
  • #5757 - Force LLVM to use ZMM registers for vectorization.
  • #5751 - [DISCUSSION] Allow CUDA array interface to be a callable

Closed PRs

  • #5768 - Interval example: fix import
  • #5764 - fix flake8 errors
  • #5759 - Support minlength argument to np.bincount

3. Next Release: Version 0.50.0, RC=June 3rd

  • 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

  • plan to have first open meeting on June 4th.

  • 0.51 potential tasks

    • TypedSet
      • base on TypedDict

      • Related:

      https://github.com/numba/numba/pull/5639
    • Nested modes
    • Move SSA pass up the pipeline
      • make sure more passes can work in SSA form
    • CUDA Dispatcher / kernel interface:
      • Share machinery with CPU target (e.g. for multiple signatures, typeconv, etc.)

4. Upcoming tasks

  • Opening up the numba meeting
    • needs to approve a CoC
    • plan to have first open meeting on week after 0.50RC
Add a custom footer
Add a custom sidebar
Clone this wiki locally