Closed jvdp1 closed 3 months ago
Thanks, @jvdp1, I'll test.
@milancurcic As discussed, using my datasets, here are the wall clock times for the steps forward-backward
and update
:
forward-backward
: 4.81supdate
: 9.44sforward-backward
: 4.80supdate
: 5.61sSo, about a 40% speedup for the step update
alone.
This PR proposes to replace some intrinsics
pack
by pointers inget_params
andget_gradients
. Pro: These changes reduced the times ofupdate
on my own tests by ~30%. Con: these procedures must beimpure