Re: [Liberty-eiffel] SoC 2016

[Bernd Schoeller]

On 06/03/16 20:40, Raphael Mack wrote:
>> https://github.com/schoelle/algae
> 
> wow, this looks like a huge effort thing - you even do the math in
> Eiffel. Of course it is always great to have the library completely in
> Eiffel, but on the other side I see the drawback, that we then really
> have to understand all the math methods - which is generally much harder
> than their use, but yes, interesting and challenging!

The main reason why I am doing this is to understand the implementation
behind all the stuff I use at work (with Python). No better way to learn
than trying to implement it yourself ;-)

But that also means that development is really slow - I am not a
mathematician and some of the numerical stuff is very hard.

My matrix multiplication in pure Eiffel is already reasonably speedy for
dense matrices: it is as fast as the numpy/blast combination that is
contained in the default python packages on Debian for my computer at home.

But it is incredible how much that can be improved by a specialized
library compiled for your machine: at work, we have a tuned version of
ATLAS that we are using for numpy, and that is rought 10x as fast as my
Eiffel solution.

My matrix inversion is using the Gaussian algorithm, which is rather
slow and mathematically not very stable.

Depending if there will be some 'real world' use for my library, I am
considering also implementing a version built on top of 'eigen'.

Bernd