Minutes_2020_06_16

Numba Meeting: 2020-06-16

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

0. Feature Discussion

• Open meeting

  • Siu: useful to hear from community, productive discussions. Good as a first meeting to hear what people were very passionate about.
  • Graham: What was said was closely aligned with pain points of which we were aware.
  • Ehsan: Scale of attendance was encouraging. Agenda/organisation could do with some improvement. Post topics ahead of time.
  • Siu: Feels similarly, first run was an experiment, now we have this experience it'll be easier to focus on a specific agenda.
  • Pearu: Agenda suggestion
    • 10 min introduce new features to Numba
  • Ehsan: Agenda suggestion
    • project direction, feedback
  • Ehsan: Level of engagement was impressive.
  • Siu: Folks essentially being representitives of larger projects.

• Next open meeting:

  • 2nd Tue of each Month
  • July 14th

• numba.discourse.group

  • Update issue template: https://github.com/numba/numba/pull/5850/files
    • Hameer signed up to review
  • TODO: put up Code of conduct

• contrib module https://numba.discourse.group/t/addons-contrib-repo/29/2

  • Hameer: concern about what is public vs private API
  • Stuart: private API needed by extensions need to move to extending. The contrib/addon package can be a guideline of what needs to be moved.
  • Ehsan: asked Stuart about numba-scipy
  • Stuart: community asked why can't scipy use numba directly
  • Hameer: scikit-learn community they numba lacks tooling to profile, debugging, etc..
  • Stuart: may be better use of resources to do that.
  • General discussion about where e.g scipy support should go.
  • Stuart/Siu prefer outside along with e.g. perhaps move NumPy.
  • Make as discussion topic for community meeting.

• Do we do a 0.50.1

  • issue: typing error in cuda are eaten
  • issue: deprecation notice didn't get bump
  • issue: get_terminal_size

• High risk items 0.51:

  • Explore moving SSA pass up the pipeline
    • test if more passes can work in SSA form
  • LLVM 10
    • optional:
      • MCJIT -> ORC-JIT ?
      • issue with MCJIT leaking JIT'ed module
      • LLVM C++ refct pruning FunctionPass
  • with objmode caching
    • risk: slowing down caching and loading "serialized" pyobject
    • https://pypi.org/project/pickle5/
  • CUDA Dispatcher / kernel interface:
    • Share machinery with CPU target (e.g. for multiple signatures, typeconv, etc.)

• Typed Set/List deprecation

  • challenging to have a switch to make typedlist by default
  • the same goes for numba.typed.Set -- once it has been written

1. New Issues

• #5868 - TypeError: compile_kernel() got an unexpected keyword argument 'boundscheck'
• #5865 - Minimum time for deprecation cycles?
• #5864 - Support for np.fft.fft, np.fft.ifft etc.
• #5863 - Add Table with llvm, llvmlite and numba compatibility to Readme
• #5860 - Typing errors in device functions aren't properly reported
  • breaking CUDA usability
• #5858 - Can numba accelerate a loop with a trained xgboost model in it?
• #5854 - Installating dependencies with python setup.py install raises SyntaxError
• #5853 - Works as plain python, core dump as @njit
• #5847 - Improved error message for non existing variable
• #5845 - Refactor deepcopy func_ir and its statements.
• #5844 - Refactor source location info into Template
• ** #5839 - Dispatching for custom types is way slower than builtin types
• #5836 - CUDA: Passing opt=0 to NVVM doesn't work
• #5835 - CUDA debug info is invalid - compile units have an empty list of subprograms
• #5831 - Make boxing two-phase to more efficiently support dependent types
• #5829 - Incorrect results when working with transposed arrays
• #5828 - native<->objmode calls overhead for small functions
• #5827 - Can't njit code containing an np.ndarray subclass

Closed Issues

• #5867 - Numba doesn't accelerate recursive function despite nopython=True passing
• #5855 - Error when install numba by pip
• #5848 - operator.pos fails on strided arrays
• #5843 - Numba 0.50.0 Checklist
• #5837 - Incorrect output when calling np.array on a list (python 3.8)
• #5832 - Slower initial compilation with newer numba from Miniconda
• #5825 - Weird behavior for loop

2. New PRs

• #5866 - [WIP] Implement str and repr builtins
• #5861 - Added except for possible Windows get_terminal_size exception
• #5857 - CUDA docs: Add notes on resetting the EMM plugin
• #5856 - Add support for conversion of inplace_binop to parfor.
• #5851 - CUDA EMM enhancements - add default get_ipc_handle implementation, skip a test conditionally
• **** #5850 - Updates the "New Issue" behaviour to better redirect users.
• #5846 - CUDA: Allow disabling NVVM optimizations, and fix debug issues
• #5841 - cleanup: Use PythonAPI.bool_from_bool in more places
• #5840 - Typed Tuple
• #5834 - Fix the is operator on Ellipsis
• #5826 - CUDA: Add function to get SASS for kernels

Closed PRs

• #5862 - Do not leak loop iteration variables into the numba.np.npyimpl namespace
• #5859 - CUDA: Fix reduce docs and style improvements
• #5852 - CUDA: Fix cuda.test()
• #5849 - Setitem for records when index is StringLiteral, including literal unroll
• #5842 - Update CHANGE_LOG for 0.50.0 final.
• #5838 - Ensure Dispatcher.__eq__ always returns a bool
• #5833 - Fixes the source location appearing incorrectly in error messages.
• #5830 - doc: Mention that caching uses pickle

3. Next Release: Version 0.51.0, RC=22 July, Final 29 July?

• Requests for 0.51

• high risk stuff for 0.51.

• 0.51 potential tasks (To be updated)

4. Upcoming tasks

• Opening up the numba meeting