[DFTB-Plus-User] GPU accelerated calculations through MAGMA solver

[Anshu Gaur]

Dear All,

I am trying to use GPU accelerated nodes on a cluster to speed-up
calculations (SCC/Optimization and transport). I have installed MAGMA
(v2.5.4) with CUDA (11.1) support. DFTB+ (20.1) is compiled with MAGMA
(config: WITH_GPU=TRUE and paths defined in intel.cmake toolchain). DFTB+
was compiled with ELSI (with internal ELPA and PEXSI solvers) using cmake (v
3.19). The compilation went smoothly and the include directory of dftb+ has
"dftbp_magmahelper.mod" present.

In my input file I have solver = MAGMA{} present in Hamiltonian. I submitted
the job on nodes with gpu acceleration. The output file contains: 
 
----------------------------------------------------------------------------
-------------------------------------------------------------------

MPI processes:               160 (split into 2 groups)
BLACS orbital grid size:     8 x 10
BLACS atom grid size:        8 x 10
Chosen random seed:          685054342
Mode:                        LBFGS relaxation
Self consistent charges:     Yes
SCC-tolerance:                 0.100000E-07
Max. scc iterations:                   1000
Shell resolved Hubbard:      No
Spin polarisation:           No
Nr. of up electrons:          1016.000000
Nr. of down electrons:        1016.000000
Periodic boundaries:         Yes
Lattice optimisation:        Yes
Pressure:                        0.000000
Electronic solver:           Divide and Conquer (MAGMA GPU version)
 Number of GPUs requested:           1
 Number of GPUs found    :         117
Mixer:                       Broyden mixer
Mixing parameter:                  0.200000
Maximal SCC-cycles:                    1000
Nr. of chrg. vec. in memory:              0
Nr. of moved atoms:                     508
Max. nr. of geometry steps:            1000
Force tolerance:               0.100000E-05

and
-------------------------------------------------------------
Using DFT-D3 dispersion corrections
Full 3rd order correction    Yes
Extra options:
                             Mulliken analysis
Force type                   original

----------------------------------------------------------------------------
----

***  Geometry step: 0, Lattice step: 0

 iSCC Total electronic   Diff electronic      SCC error
ERROR!
-> Unknown eigensolver


Am I missing something in installation/input-file or type of resources
requested?. I would really like to use MAGMA or any-other solver which uses
GPU acceleration along with CPU cores. This would specially help with
"transport" calculations where "Parallel - Groups" is limited to 1.
Any help with regards to accelerate calculations (specifically transport
calculations with > 8000 atoms in device geometry) would be really
appreciated.

Thanks,
Anshu Gaur



-- 
This email has been checked for viruses by Avast antivirus software.
https://www.avast.com/antivirus
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.zfn.uni-bremen.de/pipermail/dftb-plus-user/attachments/20210621/dff173b4/attachment.htm>