Minutes_2021_04_20
Valentin Haenel edited this page Apr 20, 2021
·
1 revision
Attendees: Todd, Graham, Guilherme, Hadia, Jim, Siu, Stu, Val
NOTE: All communication is subject to the Numba Code of Conduct.
- An update on target extension/overload work
- target overload is merged
- Graham started to adapt the CUDA target
- Stuart continue to refactor the codebase due to fallout from this work
- targets are now isolated (to some degree)
- import time gone down 40%
- Compiler flags propagation
- Concern for upcoming release
- With the amount and complexity of work we are doing, we are prioritizing stability before compiler performance. We might see some compiler speed regression in the upcoming release.
- slice None vs sys.maxsize
- https://github.com/numba/numba/pull/6939#issuecomment-822656830
- JimP: AwkwardArray picked the choice that maxsize is a magic number. So clamping the space to 2**64 - 2 numbers.
- Stu: Look at CPython unittest for known issues.
- JimP: Sparse container may use up 2**64 and more.
- JimP: without the support for bigint, operating near the limit 2**64 (or intp) is already unlikely.
- #6934 - broken link for nrt.h in "notes on numba runtime" page
- #6933 - Undocumented behaviour in prange
- #6932 - Memory leak when returning bytes from input argument
-
#6931 - Boxing for
slice
- #6930 - Inconsistent behavior of reshape between numpy and numba/cuda device array
- #6927 - Leading underscore field in structref leads to compilation failure
- #6921 - Interop With CuPy's Stream Context Mangers
- #6920 - Bad Parfor loop fusion side-effects?
- #6919 - CUDA: ffs is off-by-1
- #6925 - Possible bug when parallel=True
- #6923 - cannot import name 'add_metaclass' from 'numba.core.utils
- #6922 - cannot import name 'add_metaclass' from 'numba.core.utils'
- #6940 - Fix function resolution for intrinsics across hardware.
-
#6939 - Implement
slice
boxing - #6937 - Fix memory leak in bytes boxing
- #6936 - CUDA: Implement support for PTDS globally
- #6935 - fixes broken link in numba-runtime.rst
- #6928 - Add preliminary (limited to at most 3 dimensions) support for axis to argmax
- #6926 - [WIP] Keyword args support for Numpy functions when parallel=True option is specified
-
#6924 - CUDA: Fix
ffs