- Update to 0.53.0
* Support for Python 3.9
* Function sub-typing
* Initial support for dynamic gufuncs (i.e. from @guvectorize)
* Parallel Accelerator (@njit(parallel=True) now supports
Fortran ordered arrays
* Full release notes at
https://numba.readthedocs.io/en/0.53.0/release-notes.html
- Don't unpin-llvmlite.patch. It really need to be the correct
version.
- Refresh skip-failing-tests.patch
- Add packaging-ignore-setuptools-deprecation.patch
gh#numba/numba#6837
- Add numba-pr6851-llvm-timings.patch gh#numba/numba#6851 in order
to fix 32-bit issues gh#numba/numba#6832
- Update to 0.52.0
https://numba.readthedocs.io/en/stable/release-notes.html
This release focuses on performance improvements, but also adds
some new features and contains numerous bug fixes and stability
improvements.
Highlights of core performance improvements include:
* Intel kindly sponsored research and development into producing
a new reference count pruning pass. This pass operates at the
LLVM level and can prune a number of common reference counting
patterns. This will improve performance for two primary
reasons:
- There will be less pressure on the atomic locks used to do
the reference counting.
- Removal of reference counting operations permits more
inlining and the optimisation passes can in general do more
with what is present.
(Siu Kwan Lam).
* Intel also sponsored work to improve the performance of the
numba.typed.List container, particularly in the case of
__getitem__ and iteration (Stuart Archibald).
* Superword-level parallelism vectorization is now switched on
and the optimisation pipeline has been lightly analysed and
tuned so as to be able to vectorize more and more often
(Stuart Archibald).
Highlights of core feature changes include:
* The inspect_cfg method on the JIT dispatcher object has been
significantly enhanced and now includes highlighted output and
interleaved line markers and Python source (Stuart Archibald).
* The BSD operating system is now unofficially supported (Stuart
Archibald).
* Numerous features/functionality improvements to NumPy support,
including support for:
- np.asfarray (Guilherme Leobas)
- “subtyping” in record arrays (Lucio Fernandez-Arjona)
- np.split and np.array_split (Isaac Virshup)
- operator.contains with ndarray (@mugoh).
- np.asarray_chkfinite (Rishabh Varshney).
- NumPy 1.19 (Stuart Archibald).
- the ndarray allocators, empty, ones and zeros, accepting a
dtype specified as a string literal (Stuart Archibald).
* Booleans are now supported as literal types (Alexey Kozlov).
* On the CUDA target:
* CUDA 9.0 is now the minimum supported version (Graham Markall).
* Support for Unified Memory has been added (Max Katz).
* Kernel launch overhead is reduced (Graham Markall).
* Cudasim support for mapped array, memcopies and memset has
been * added (Mike Williams).
* Access has been wired in to all libdevice functions (Graham
Markall).
* Additional CUDA atomic operations have been added (Michae
Collison).
* Additional math library functions (frexp, ldexp, isfinite)
(Zhihao * Yuan).
* Support for power on complex numbers (Graham Markall).
Deprecations to note:
* There are no new deprecations. However, note that
“compatibility” mode, which was added some 40 releases ago to
help transition from 0.11 to 0.12+, has been removed! Also,
the shim to permit the import of jitclass from Numba’s top
level namespace has now been removed as per the deprecation
schedule.
- NEP 29: Skip python36 build. Python 3.6 is dropped by NumPy 1.20
OBS-URL: https://build.opensuse.org/request/show/880602
OBS-URL: https://build.opensuse.org/package/show/devel:languages:python:numeric/python-numba?expand=0&rev=47
Dirk Mueller
ef1752b2cb