Minutes_2020_12_15

Numba Meeting: 2020-12-15

Attendees: Siu, Dale, Michael, Stuart, Todd, Val, Ehsan, Guilherme, Hameer

Last meeting of the year!
Open meetings
- Jan 5th 2021 first open meeting
- Zoom
- More on weekyl themes/topics later
Python 3.9 update
- changes in bytecode, C struct/API
- has open PR https://github.com/numba/numba/pull/6579
Python 3.6 EOL (security support stops in Dec 2021).
- maybe follow numpy's plan
- needs announce if we are dropping py version. earliest to drop is Numba-0.54
Expose LLVM optimisation levels
- 0.52 fixed optimization satbility issue. Details: a function optimizes differently depending on whether it is inlined into the cpython wrapper or not.
- Because of that, some code has a different performance characteristic because the optimization pipeline changed.
- How to?
  - Per-decorator changes are likely to be asking for trouble
  - Global switches will be easier
- Dicussions:
  - Per-decorator opt-level allows for manual fine-tuning; i.e. a lot of functions are plumbing and few are numeric.
  - Per-decorator option will allow for opt-out from expensive passes.
  - might need a "hint" module to help marking code.
  - few projects need to opt for size; but there projects deploying to phones as well.

#6576 - Unexpected "divide by zero" warning with numba.vectorize and np.log
- LLVM opt
#6572 - Add Zenodo badge
- Not sure about benefit and value
- Is the maintenance burden satisfieable
- Ask on Discourse?
- Pandas and Numpy already doing this
#6571 - All Numba internal registrations should happen before extensions are run.
- This will become a discussion
**** #6570 - Dicts with tuple keys not working in 0.52.0
- Casting?
- Code maybe illegal, typecast reckons it is safe, but probably it isn't as the values/types are being used for lookup
#6569 - numba.core import issues
- problem installing numba w/o internet access or can't use prebuilt binaries
#6564 - Inconsistant result between numba and py_func
- Age old request for pairwise sum
#6563 - CUDA: Add support for passing tuples to kernels via vectorize
- Can be a feature request
#6561 - Cuda runtime errors when starting up numba
- ISA mismatch on PTX, installation probbably broken
#6559 - Migrating to LLVM 11 (on 64bit RISC-V)
#6558 - Wheel support for linux aarch64[arm64]
- There are containers now, so we can do this now
#6557 - Separate the LLVM code correction in nvvm.py into a separate function
- PR to implement this exists already, see below
#6556 - Error in compiled function using CFFI to call BLAS
- Things being free'd too early
- Might have a good solution, adding increfs
- Maybe try moving decref to the end of basic blocks in Numba IR

#6575 - Avoid temp variable assignments
- *** Maybe an optimization pass
- pending on 3.9
#6574 - Run parfor fusion if 2 or more parfors
- Is fine, just needs a small fix
#6573 - Improve __str__ for typed.List when invoked from IPython shell
- Maybe needs some more work to determine if this is the correct way to do this
- Why flake8 not failing?
#6567 - Refactor llvm replacement code into separate function
#6560 - Add LLVM pass timer