[Q-e-developers] LOBPCG, a new diagonalization algorithm in Quantum ESPRESSO?

[Andrew Knyazev]

Dear Paolo Giannozzi,

I have looked at the Quantum ESPRESSO code docs. It appears that the 
code has only the  standard diagonalization algorithms: CG and Davidson. 
You may be interested to try implementing a new one: LOBPCG, see
http://en.wikipedia.org/wiki/LOBPCG

It is currently the default algorithm for parallel/GPU runs in ABINIT, see
http://www.abinit.org/documentation/helpfiles/for-v7.0/input_variables/vardev.html#wfoptalg
http://www.abinit.org/documentation/helpfiles/for-v7.0/input_variables/vardev.html#gpu_linalg_limit

The implementation is described in
http://www.sciencedirect.com/science/article/pii/S0927025607002091
The code has actually been written by my collaborators in this paper, so 
please ask them if you have specific questions on implementation.

Compared to the standard diagonalization algorithms, LOBPCG is typically 
faster and more robust, but, most important.y, easy to parallelize.

Due to the nature of my current position, I cannot contribute to code 
development,  but I would be glad to help with an advise, if you are 
interested.

-- 
Regards,

Andrew