Re: Numba for Debian Science
- To: debian-science@lists.debian.org
- Subject: Re: Numba for Debian Science
- From: Daniel Stender <debian@danielstender.com>
- Date: Wed, 26 Jul 2017 18:43:22 +0200
- Message-id: <[🔎] 11aa5ff6-c575-b0bb-fb84-c5ad4c8c4c0a@danielstender.com>
- In-reply-to: <d887413a-1cf2-25be-1265-86226cac1429@danielstender.com>
- References: <1482003743.1472.2.camel@gmx.com> <bc4c4339-a848-9467-eb10-55d4934ef7ca@danielstender.com> <e8d2935a-7814-987f-6ecb-b5293c69d634@danielstender.com> <e11ca6fb-3f05-77a0-a725-d861afee1b71@danielstender.com> <d887413a-1cf2-25be-1265-86226cac1429@danielstender.com>
On 07.06.2017 18:24, Daniel Stender wrote:
> On 27.05.2017 23:06, Daniel Stender wrote:
>> There some progress on Numba. I've got 0.33.0 in experimental now, with llvmlite 0.18.0 running
>> on LLVM 4.0. Sorry for the delay.
>>
>> The good news is, the problem with llvm.get_host_cpu_features() yielding a runtime error on
>> some archs disappeared, and Numba is ready for Numpy 1.12 now.
>>
>> Unfortunately, it remains being a test failure fest, and builds nowhere w/o problems:
>> https://bugs.debian.org/cgi-bin/pkgreport.cgi?repeatmerged=no&src=numba
>>
>> I've posted bug reports on the individual issues, if you want to fix some and provide patches,
>> welcome.
>>
>> DS
>
> Due to the lack of enough spare time for fixing this to get into Unstable in the near
> future, I've set Numba to RFA (#864367) now.
>
> Thanks,
> Daniel Stender
Note: Numba [1] 0.34.0 is in unstable now. Although the package still needs a couple of fixes to go into
testing, on amd64 and some other archs it appears to run properly.
Nutshell: Numba is an on-the-fly/just-in-time (JIT) machine code compiler for Python. Just decorate
certain computation intensive functions with "@jit" to speed them up and execute the script, that's
basically it [2]. It's pretty straightforward to use. Under the hood, LLVM resp. llvmlite [3] is used
to perform the compilation into native machine code.
/usr/share/doc/numba-doc/examples contains a couple of examples:
$ ./bubblesort.py
== Test Pure-Python ==
[0 1 2 3 4 5 6 7]
== Test Numba ==
[0 1 2 3 4 5 6 7]
Python 0.257357192039
Numba 0.000370597839355
Speedup 694.437918168 x [Intel Core i5-6200U]
Numba especially integrates with the Python scientific stack and supports accelerated Numpy array computation
on CUDA/GPU devices, I'll write more about that on Planet Debian, soon.
Thanks,
Daniel Stender
[1] https://packages.qa.debian.org/n/numba.html
[2] http://numba.pydata.org/numba-doc/latest/index.html
[3] https://packages.qa.debian.org/l/llvmlite.html
--
4096R/DF5182C8 (stender@debian.org)
http://www.danielstender.com/
Reply to: