sort_by_key - new algorithm request about compute HOT 4 CLOSED

Sounds great! The sort_by_key() algorithm is definitely on my big TODO list (I just need to find some spare time). Can you let me know what key/value data types you would be using?

from compute.

My code implements object properties as multiple arrays of a single property (to allow efficient access on GPU's). The most common value data type would be an integer (32-bit) that would index the various property arrays.

Key types can be a variety of things. Some examples of cases that I have are:

Simple cases:

float - the simplest case
floatN - ND data, sort by a single coordinate only (selectable at run-time)

More complex:

float2 - 2D data, sort by X then Y
floatN - ND data, sort by multiple coordinates in a specified order (i.e. first X, then Z then Y)

Of course integer and double key types would be very nice to have as well. :-)

from compute.

Hi George,

I've added initial support for the sort_by_key() algorithm (commit is here 3389a5c). As of right now it's only implemented for the serial insertion sort back-end so it is quite slow. But for now you can try compiling it and sorting small sequences.

Thanks,
Kyle

from compute.

I've updated the sort_by_key() algorithm to use radix sort instead of insertion sort (see commit bae7432). Performance should now be much better.

from compute.