Pythran version

[serge-sans-paille]

No description provided.

[MasonProtter]

Nice! Is there a way to make this work with user-defined types? Ie. the equivalent of the Julia example with an Array of Point objects?

[SimonDanisch]

I'm guessing that this is work in progress, since there is still the actual n-dimensional, lazy broadcast implementation missing ;)

[serge-sans-paille]

lazy broadcast implementation missing

It's all implemented within pythran, Is there any meaning of reproducing it here?

[serge-sans-paille]

Ie. the equivalent of the Julia example with an Array of Point objects?

It works for any shape of ndarray, so I guess the answer is « only if your reprensent your array of Point as an ndarray. »

[SimonDanisch]

It's all implemented within pythran, Is there any meaning of reproducing it here?

yes, the one and only point of this challenge is to evaluate, how the implementation looks and how much effort it is to make it nice and see how flexible it is in its core - it's all to better evaluate how well the implementation can grow in the future and how easily new contributors can start working on it. I think there is a lot of development currently happening in machine learning, and flexible and easily maintainable array algorithms will be crucial to tackle that challenge

[MasonProtter]

It works for any shape of ndarray, so I guess the answer is « only if your reprensent your array of Point as an ndarray. »

Part of what makes Julia so wonderful (and I believe one of the things Simon was trying to point out with this challenge) is that Julia is nearly unique[1] in how easy it is to generalize an algorithm to work on any datatype without doing any violence to that datatype. Like for instance, sometimes one wants to broadcast over arrays of strings or arrays of dual numbers or arrays of a composite datatype containing a string, a float64, two different integers, a symbol and a matrix.

Because package designers can't always predict what their types and algorithms will be used for, it's important that they are maximally flexible. If you're looking for generality, ndarrays of Float64s are pretty far from optimally general.

[1] Though the Nim solution really impressed me! Clearly, such things are possible in Nim too (hence the 'nearly'.

[oxinabox]

Part of what makes Julia so wonderful (and I believe one of the things Simon was trying to point out with this challenge) is that Julia is nearly unique[1] in how easy it is to generalize an algorithm to work on any datatype without doing any violence to that datatype

Or rather: Julia programmers in general think that Julia is unique in this way.
And this challage is a chance to prove us wrong.

(I hope we are wrong, because I want thus awesome flexible code everywhere)