Minutes_2022_10_04
Siu Kwan Lam edited this page Oct 4, 2022
·
1 revision
Attendees: Jim Pivarski, Ianna Osborne, Val, Siu Kwan Lam, Benjamin Graham, Luk, Todd A Anderson, stuart, kc611, Guilherme, Graham, Markall, LI Da, Sebastian Berg, Andre Masella, FPOC (last week): Stuart
NOTE: All communication is subject to the Numba Code of Conduct.
Please refer to this calendar for the next meeting date.
- Deprecate PYCC (Stu)
- PYCC has multiple longstanding issues and upcoming problems with setuptools; it also has very limited feature
- Target to deprecate in 0.57 release
- No one in the meeting objected
- numpy
ufunc.typeschanged; contains extra SIMD loops #8478 (Stu)- What API would actually help numba here?
-
np.ufunc.resolve_dtypes(dtypes)?- this would return dtypes used by the inner loop.
- Do you have need for
np.ufunc.get_loop()(to get the inner loop)?- useful for complicated inner loop so we don't need to duplicate
- Do you prefer C-API, or Python API?
- both are useful.
- Prioritizing Python is fine from Numba p.o.v.
- Dufunc can be implemented as a promotion-step within ufunc. The promotion can insert new inner loop. (This is in development in NumPy)
- Advanced performance tricks to go into the docs (Luk, Stu)
- Initial step, to start a issue to collect ideas before adding to the docs
- Release 0.56.3 plan:
- unpin setuptools https://github.com/numba/numba/pull/8475
- Fix #8477: Allow copies with different strides for 0-length data https://github.com/numba/numba/pull/8482
- #8470 - Deconstructor support in jitclass
-
#8471 - Lack of support for
setitemsimultaneously across all fields of record arrays -
#8472 - The
.fillon a record array results in aKeyError - #8477 - CUDA: Copying zero-size device array to zero-size host array can fail if strides differ
- #8478 - fmin and fmax incorrect output dtype
- #8479 - Support type change across inplace operators
- #8480 - TypingError: Use of unsupported NumPy function 'numpy.cbrt' or unsupported use of the function.
- #8483 - CUDA: Tuples of CAI exporters that aren't Numba device arrays are not recognized by typing
- #8485 - Feature Request: Support math.fsum
- #8481 - ImportError: Numba requires at least version 0.40.0 of llvmlite.
- #8468 - Reference test suite in installation documentation
- #8469 - Correctly handle optional types in parfors lowering
- #8473 - change the include style in _pymodule.h and remove unused or duplicate headers in two header files
-
#8475 - Remove setuptools version pin
- Siu to make PR to skip pycc tests due to failure of incompatible VC in buildfarm. We are choosing this fix for pragmatic reason.
- #8476 - Make setuptools optional at runtime.
- #8482 - Fix #8477: Allow copies with different strides for 0-length data
- #8484 - Initial commit for raise/assert in device functions
- #8486 - Restrict the TBB development package to supported version in Azure.
- #8474 - Remove setuptools as a runtime dependency