GROMACS development

03/29/2019

02:29 PM GROMACS Bug #2641: Possible l-bfgs improvements

Ahh, OK seems like I've used a wrong branch.
Here is your minimize.cpp file from the patch on Gerrit.
I've just add...

12:37 PM GROMACS Bug #2641: Possible l-bfgs improvements

I've made L-BFGS in master branch to behave like an old 4.6.5 (up to some numerical precision)
minimize.cpp with cha...

03/27/2019

09:16 AM GROMACS Bug #2641: Possible l-bfgs improvements

Yes, That's why I've spent last evening to check other implementations of L-BFGS in the following library:
https://g...

03/26/2019

05:01 PM GROMACS Bug #2641: Possible l-bfgs improvements

OK, I agree that this is just semantics.
Lets then normalize s[] with ||s|| = snorm.
Lets also define smax as a...

01:16 PM GROMACS Bug #2641: Possible l-bfgs improvements

OK, I've looked through the whole code again, so
1) We have direction of search given by the array s[] which is al...

12:18 PM GROMACS Bug #2641: Possible l-bfgs improvements

Yes exactly, I've also noticed that it affects only double precision, because line search never reaches 20 steps in s...

10:53 AM GROMACS Bug #2641: Possible l-bfgs improvements

Hi I've found another bug in the minimize.cpp: line 2166
It is now:
if (std::fabs(sb->epot - Epot0) < G...

09:54 AM GROMACS Bug #2641: Possible l-bfgs improvements

Why there are such a strange things like:
stepsize = 1.0/fnorm;
in the Gromacs l-bfgs code?
I mean stepsize is ...

03/25/2019

11:27 AM GROMACS Bug #2641: Possible l-bfgs improvements

Hi,
I've identified the problem with current L-BFGS implementation.
Seems like it has been broken at some point c...