stable reversed indirect sort?

[rainwoodman]

Looks like there is no easy way to achieve a stable reversed indirect sort with the sorting functions in numpy:

argsort()[::-1] would give a reversed indirect sort.

argsort(kind='mergesort')[::-1] does not give a stable reversed indirect sort. ([::-1] makes it unstable.

Would you consider including a 'reversed' parameter to argsort?

[rainwoodman]

The easiest API change is to allow

argsort(order=[('field0', 1), ('field1', -1), ('field2', -1)])

Then we can make use of the new NPY_AUX_DATA to pass in such flags for reversed comparison used by 'compare' function.

[jaimefrio]

It would be much easier to implement a two-pass O(n) stable reversal of the indexing array as a postprocessing step. I'm not sure we can do what you suggest without significantly affecting performance: sorting moved away from using the dtype's compare function, to having specialized inlined functions with hardcoded < for native types. And I don't think we can pass a "pointer to an inline function" and get the same performance we have now...

[jaimefrio]

So I guess I'm leaning towards a reverse kwarg solution. Before getting something like this going it would be a good idea to talk about it on the mailing list. Could you send an e-mail to get the discussion started?