Minutes_2020_03_17

Numba Meeting: 2020-03-17

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

SSA and pending issues will likely delay release by 1 week
- current status on SSA:
  - mainly blocked by 3 parfors tests failure
  - https://github.com/numba/numba/pull/5351#issuecomment-600134414
- other issues must go in
  - parfor gufunc tuple PR
  - public shim PR
First-class function type
- 2 issues with the PR
  - caching not working first-class function type
  - more than 1 signature may have typing conflict on the return-type
TODO:
- check cuda testsuite on our DGX for segfault

#5391 - the right way to jit a matrix dot function (why so slow?)
- question
#5389 - NotImplementedError: only one advanced index supported
- there's workaround
#5387 - Prevent LLVM parsing errors arising from nonsensical code and literal_unroll
- needs better error checking in numba and llvmlite
#5385 - np.cbrt cannot support in nopython mode
- feature request
- should be fine now (C99)
#5383 - Crash with parallel=True for array shapes computed via tuple concatenation
- Parfor not handling the case
- ConstProp would help
#5380 - Error raised when using functools.reduce() together with operator.mul, and setting nopython=True
- support request
- and need more reproducers
#5379 - In wheels bin/numba is not executable
- odd wheel problem
- need to fix before release
#5377 - Confirm support for try/except use in objmode block
** #5375 - Deprecate record access by attr in CUDA target
- we can start deprecation process this release
- More likely next release
#5374 - Shared Memory behaves inconsistently in NUMBA_ENABLE_CUDASIM
#5373 - Numba Sort is Significantly Slower Than NumPy Sort
#5372 - Produce better error message for the unsupported case of iteration over an empty tuple.
- empty tuple should be unituple
#5371 - 0.48.0 breaks TBB support (installing from pip)