While working on more radical FFM optimization, I've noticed things were done in a very convoluted way with a needless temporary array involved.
Removal of that array and refactoring of code related to it leads to about 10% speed improvement of FFM part of the code in my tests both on server and on laptop.
This only applies to training code. No speedup in pure prediction mode.
While working on more radical FFM optimization, I've noticed things were done in a very convoluted way with a needless temporary array involved.
Removal of that array and refactoring of code related to it leads to about 10% speed improvement of FFM part of the code in my tests both on server and on laptop.
This only applies to training code. No speedup in pure prediction mode.