Minutes_2020_04_21

Numba Meeting: 2020-04-21

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

0. Feature Discussion

• Move examples into github.com/numba/numba-examples
  • like np.einsum
  • lower barrier to submit examples
  • easier for user community to contribute
  • easier to test using our integration test
• subtyping proposals
  • record subtyping (row polymorphism)
    • possible extension to tuple/namedtuple; but it can't be pass-by-value
    • is it safe? would broken by branch-pruning?
    • sensitive to ordering
  • TODO next week:
    • function subtyping
    • and all the above concerns
• 0.49.1?
  • likely need the bugfix release because it's the last release to retain the old API
  • a few regression caused likely by SSA
  • wait until next week for issues to be reported
• and llvmlite 0.32.1
  • invalid pypi tags
• Q&A forum that's not gitter or mailing list
  • gitter:
    • high traffic, hard to search, threads in beta
    • we should report issues
  • StackOverflow:
    • low signal to noise ratio
    • may have improved; "numba" tag working
  • mailing list:
    • low traffic, old
  • Discourse
    • may need investigation
  • TODO next week
    • try observing SO "numba" tag for a little

1. New Issues

• #5590 - cannot determine Numba type of <class 'lmfit.parameter.Parameter'>. Error while using Numba with lmfit
  • closed
• #5589 - Internal error "module 'numba' has no attribute 'core'" in Python 3.6
  • circular import accident
  • mixed env?
  • pip related?
• **** #5588 - Make _get_thread_id() a public
  • threading backend dependent
  • thread-id function assumes single thread team?
  • potential for misuse
  • can be solved by higher-level features; i.e. atomic operations
• #5586 - Incorrect Directory Reference to numba/npyunfunc
  • need to check all docs for bad refs
  • patched in #5587
• #5584 - BUG: Bug with parfors reduction (parallel=True) in 0.49
  • very likely caused by SSA
• #5583 - cannot import name 'cffi_support' from 'numba'
  • need to check all examples for bad refs
  • patched in #5592
• #5582 - Support generic attribute types for jitclass
• #5577 - numba.roc.atomic erroneously assumes ints
  • Need to move to LLVM 9
  • Move roctools to llvmlite
• #5576 - Invalid __global__ write with dynamic shared memory, indexing by block index, and a for loop over a range
• #5575 - A typo affecting binop typeinfer in parfor's _arrayexpr_tree_to_ir
  • easy to fix
  • parfor bug
• #5570 - Datashader/Numba various LoweringErrors
  • likely SSA related
• #5566 - RuntimeError: cannot cache function '__shear_dense'
  • root user and cache
  • librosa
• #5562 - No performance improvement with Intel's SVML library
  • TODO: remove all remaining win32 math shim to see if it fixes this.
• #5559 - Part of compilation failure message is dropped when handling literal args
• **** #5554 - [For discussion] Subtyping proposal
• #5552 - Declarative Typing in overloads
• #5551 - Function does not stop at SIGINT

Closed Issues

• #5585 - builtins.ImportError: Numba could not be imported.
• #5565 - literally fails when getitem by string literal index is involved
• #5564 - Numba 0.49.0 Checklist
• #5563 - Environment variables with NUMBAPRO prefix are depreciated
• #5550 - No function '__pyx_fuse_0pdtr' found in __pyx_capi__ of 'scipy.special.cython_special'

2. New PRs

• #5593 - remove decref in pop
• #5592 - Fixes #5583 Remove references to cffi_support from docs and examples
• #5591 - Fix #5525
• #5587 - Fixed #5586 Threading Implementation Typos
• #5579 - [WIP] Function subtyping
• #5578 - Implement math.modf for CUDA target
• #5573 - CUDA: Skip tests under cuda-memcheck that hang it
• #5561 - [WIP] CUDA: Add function for compiling to PTX (+ other small fixes)
• #5560 - Record subtyping
• #5558 - CUDA: Only release primary ctx if retained
• #5556 - CUDA: Make atomic semantics match Python / NumPy, and fix #5458

Closed PRs

• #5581 - Revert #5568
• #5580 - Revert shim
• #5574 - Update CHANGELOG for 0.49 wrt final patches
• #5572 - [DON'T MERGE] Pin down llvmlite requirement
• #5571 - Pin down llvmlite requirement
• #5569 - Fix RewriteConstGetitems not registering calltype for new expr
• #5568 - Add skip to entrypoint test on windows
• #5567 - Fixes to the change log post 0.49.0rc2
• #5557 - First-class function call cannot use keyword arguments
• #5555 - Fix #5515 by reverting changes to ArrayAnalysis
• #5553 - numba.postproc moves to numba.core.postproc

3. Next Release: Version 0.50.0, RC=May 28th

• 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