Minutes_2021_03_16
luk-f-a edited this page Mar 17, 2021
·
2 revisions
Attendees: kalyan, Ehsan, Todd, Jim P., Graham, Stuart, Siu, Val, Guilherme
NOTE: All communication is subject to the Numba Code of Conduct.
- 0.53.0 Regression
- dynamic gufunc pickling error
- cuda regression
- cuda array interface doesn't work w/ mapped arrays
- Follow up on Mentorship program
- draft: https://hackmd.io/I364JFMkSYmA_T3qwZyeSQ
- Q:
how long should the program last?how often to review program? 4 weeks - Q: minimal commitment: 3 mos
- Q: do we need changes to the current triage rotation?
- pairing rotation gives mentees more experience
- Q: max number participants: ~3
- once graduated:
- participants become long term triager
- gains GH triager rights(?)
- criteria:
- commitment on long term contribution
- FP flags not modelled in LLVM
- 4-5 issues
- make FAQ
- Related LLVM thread:
-
#6825 - reflected list behaves differently on Python 3.7 and 3.8
- likely array comprehension bug
-
#6824 - Array contains non-contiguous buffer error even after using np.ascontiguousarray
- legit bug
-
#6821 - REGRESSION: guvectorized functions can no longer be pickled in numba 0.53
- will need a 0.53.1 release
- fix is probably quite small
-
#6819 - RuntimeWarning: invalid value when disabling AVX2
- error-model not propagated
- LLVM optimization moves instruction ignoring FP-flags
-
#6812 - NumPy 1.20 numerical regressions
- caused by a change in np1.20
- #6823 - REGRESSIOIN: LoweringError when importing pynndescent
- #6820 - declaring the type for numba jit is slower than not?
- #6815 - If condition against an integer fails
-
#6822 - [WIP] Add support for np.rot90
- Also being discussed on: https://numba.discourse.group/t/use-np-transpose-with-axes-argument-for-arbitrary-array-dimension/612/4
- get's into some tricky parts of Numba
- #6818 - CUDA: Support IPC on Windows
- #6817 - Update changelog for 0.53.0 final release date
- #6814 - Improving context stack
-
#6813 - Mostly CUDA: Replace llvmpy API usage with llvmlite APIs
- Needs to be
icmp_signed
oricmp_unsigned
- Fails only for CUDA 9.0 10.0
- Needs to be