[Pw_forum] Calculation Precision on QE5.0.2-GPU

Filippo Spiga spiga.filippo at gmail.com
Fri Feb 21 10:47:58 CET 2014


Dear Tobin,

you problem is not new, other people (in China) pointed to me about it. So the issue you see it is 99% due to a possible bug in ADDUSDENS. I have been able to reproduce it locally on my cluster.  Please ad the flag "-D__DISABLE_CUDA_ADDUSDENS" to make.sys, it will make the issue disappear. I am investigating during these days to find a fix and I will commit it as fast as I can.

The ML list for QE-GPU is q-e-gpgpu at qe-forge.org Please send there any queries concerning the package, I rarely follow what is posted on the main PWscf mailing-list.

Good luck for the competition
F


On Feb 21, 2014, at 3:56 AM, Tobin Chen <chenzhao.me at gmail.com> wrote:
> hi everyone:
>     It's my first time to use QE, and I meet some problems that I don't know how to handle it.
>     DESCRIPTION
>     At first, I compile the QE5.0.2 without GPU. And I use pw.x to run my relax.in(attach at the end), using the command:
>         [PATH_TO_QE]/bin/pw.x -in relax.in
>     By this version, I can run relax.in to the end.
>     However, when I recompile QE with Nvidia GPU (QE-GPU-r216.tar.gz with the QE-5.0.2_GPU-r216.patch, and I use CUDA5.5), I meet the one of the Frequent errors during execution, which is 
>         %%%%%%%
>             Error in routine electrons (1):
>             charge is wrong
>         %%%%%%%
>     So I change the input file by comment some settings on SYSTEM modules.
>         !  occupations='smearing',smearing='gaussian',degauss=0.002,
>         !  nspin=2
>     And then I can run it on my gpu host, using my gpu card. For this situation, I guess the "gaussian smearing" may lost some precision on GPU calculation. But I'm not sure that.
> 
>     QUESTION
>     I feel confused of this situation. My question is
>     1, Is the GPU's precision not support for  some of the QE workload? or just some algorithms?
>     2, If it's just for the algorithms problem, which algorithm would be affect during calculating on GPU version? How should I do when I want to use GPU to accelera the QE-CPU version?
>     3, The last question is, what changes would happen to the final result when I modify the parameter "occupations" and "smearing"? Just like the PW/tests/check-pw.x.j, if I don't have the reference, How can I check whether my result is true when I calculate for the new workload?
> 
>     Thanks for everyone ! Maybe my chinglish would confuse you, but please report it to me. Because I really want you help !
>     Thank you very much !
> 
> -- 
> Tobin Chen
> Sun Yat-Sen University, Guangdong, China.
> 
> file relax.in
> &control
>    calculation = 'vc-relax'
>    prefix='Na2Fe2As2O',
>    pseudo_dir ='/home/jin/projects/asc14/suanli/test2/workload1/workload2-upf'
>    outdir='./tmp'
>         etot_conv_thr = 1.0E-5 ,
>         forc_conv_thr = 1.0D-4
>    tprnfor=.TRUE.
>    disk_io='none'
> /
> &system
>    ibrav=7,
>    celldm(1) =7.691188393, celldm(3)=3.756265356,
>    nat=7, ntyp=5,
>    ecutwfc=40, ecutrho=480,
>    occupations='smearing',smearing='gaussian',degauss=0.002,
>    nspin=2
>    starting_magnetization(2)=0.125
>    starting_magnetization(3)=-0.125
> !  nbnd=35
> !  lda_plus_u=.TRUE.
> !  Hubbard_U(2)=6.0
> /
> &electrons
>    electron_maxstep=300
>    mixing_beta = 0.3
>    conv_thr =  1.0d-10
> /
> &ions
>   bfgs_ndim= 3,
>   ion_dynamics='bfgs'
>   pot_extrapolation = 'second_order' ,
>   wfc_extrapolation = 'second_order' 
> /
> &CELL
>   cell_dynamics = 'bfgs' 
> /
> ATOMIC_SPECIES
> Na   22.99  Na.pw91-sp-van_ak.UPF 
> Fe1  55.845 Fe.pw91-sp-van_ak.UPF 
> Fe2  55.845 Fe.pw91-sp-van_ak.UPF 
> As   74.92  As.pw91-n-van.UPF
> O     16.00  O.pw91-van_ak.UPF 
> ATOMIC_POSITIONS {angstrom}      
> Na      0.0000000000     0.0000000000     4.7989032000
> Na      0.0000000000     0.0000000000    10.4890968000
> Fe1     0.0000000000    -2.0350000000     7.6440000000
> Fe2     2.0350000000     0.0000000000     7.6440000000
> As      0.0000000000     0.0000000000     1.8529056000
> As      0.0000000000     0.0000000000    13.4350944000
> O       0.0000000000     0.0000000000     7.6440000000
> K_POINTS {automatic}
> 4 4 4 1 1 1
> npool=2
> 
> <relax.in>_______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://pwscf.org/mailman/listinfo/pw_forum

--
Mr. Filippo SPIGA, M.Sc.
http://www.linkedin.com/in/filippospiga ~ skype: filippo.spiga

«Nobody will drive us out of Cantor's paradise.» ~ David Hilbert

*****
Disclaimer: "Please note this message and any attachments are CONFIDENTIAL and may be privileged or otherwise protected from disclosure. The contents are not to be disclosed to anyone other than the addressee. Unauthorized recipients are requested to preserve this confidentiality and to advise the sender immediately of any error in transmission."






More information about the users mailing list