[ANN] Numba 0.56.0

Dear all,

on behalf of the Numba crew I am happy to announce the availability of Numba 0.56.0 final.

As usual, you can install this release with conda from the numba channel on anaconda.org:

conda install -c numba numba=0.56.0

Or alternatively install the wheels form PyPi with pip:

pip install numba==0.56.0

You can find the release notes at: Release Notes

Thank you all!

6 Likes

Iā€™m not sure if this is the right place to post this. I tried Miniconda today. By default, the tbb package is missing unlike the full Anaconda installation.

conda install tbb
conda install -c numba numba=0.56.0

For the wheels installation via pip, I came to realization that Numba does not work with tbb/tbb-devel 2021.6.0 due to installing the latest and building Numba manually.

pip install tbb==2021.5.0   # macOS
pip install tbb==2021.5.1   # Linux
pip install tbb==2021.5.2   # Windows
pip install numba==0.56.0

Thanks to the numba team for this release.

I would like to shed some more light on a point mentioned in the release notes:

  • The performance of @stencil s has been improved in both serial and parallel execution.

As mentioned the performance of @stencil has definitely improved in both serial and parallel mode. I reran the tests for Issue#6456 that I first reported and here are the results:

10 runs (M1 chip)                           min.     avg.                        
njit + stencil:                            2.436s   3.012s
Parallel njit + stencil:                   1.718s   1.838s
njit + Loops:                              4.617s   4.694s
Parallel njit + Loops:                     3.685s   3.791s
Explicit Parallel njit + Loops + prange:   2.608s   2.727s

Special thanks to @DrTodd13 for the pull request.

2 Likes