[SciPy-User] Should one NOT start with Cython first?

[Matthew Brett]

Hi,

On Fri, Mar 13, 2015 at 10:25 PM, Scott High <highscot at gmail.com> wrote:
> Hey All,
>
> First:
>>
>> I have promised myself to never start writing "Cython-first" in mind
>> again.
>
> I agree, good choice.
>
> Beyond that, I have a few comments based on my experience using
> Python/Cython for research in numerical methods.
>
>> - Most code in scientific programming is not CPU bound. You don't get a
>> faster harddrive or faster network connection from using Cython. This
>> accounts for the majority (often 80 to 90 %) of the code we write.
>
> This is often true, but is not necessarily an argument again using Cython.
> (You only mentioned hard drives and network connections, but hitting any
> level in the memory hierarchy that is not cache can wreck performance). If
> you carefully read code that uses numpy array operations (especially
> slicing) you will notice that data will often be accessed multiple times in
> the same mathematical operation. The largest speed ups I have gotten with
> Cython are not from reducing the number of flops, but in using C style loops
> to minimize the number of passes through memory. For the computer scientists
> in the room: Cython can often allow you to better utilize cache. For me a
> good rule has been: usually don't convert to Cython to reduce the number of
> flops, but consider converting to better manage memory. There are some
> serious weasel words in that rule, mostly because it is very hard to judge
> potential speed ups by inspecting code, even for experts.
>
>> They key here is that Cython is really great, and can be very easy to use
>> -- but it is not at all cost free. The edit-compile-test cycle is much
>> slower, and it's a lot harder to debug than Python -- so you really don't
>> want to use it unless there is a real gain to to doing so.
>
> If you intend to write the majority of your program in Cython the compile
> time can really hurt you, and I am not crazy about the large amount of
> unreadable code generated. On the other hand, I have found that if you only
> intend to convert a small number of functions then the IPython Cython cell
> magic can make development nearly as quick as it would be with pure Python.
> As far as debugging goes, it is a pain. No way around that one.
>
>> - Optimizing NumPy code with Numba is often easier than using Cython.
>
> I have used Numba and was very much impressed. The problem is that
> installation can be a nightmare. Depending on your users system they may
> have to install LLVM (I had to build it from source). This is a very heavy
> duty dependency, do not discount it.

Yes, I would think hard before committing to numba.  Maybe because of
dependencies, or some other reason, I believe it is fair to say that
Cython is in much wider use than numba.  For example, I don't know if
any use of numba in the packages I use (or write) but Cython is very
common.

Cheers,

Matthew