The OpenMX1 is a numerical simulation code based on density functional theory, and is developed by Dr. Taisuke OZAKI and OpenMX developpers.
This patch is a set of codes tuned for speeding up computation on vector processors. Here, the vector processor is assumed to be NEC SX-Aurora TSUBASA, and is a processor with a very long vector length (256 or more). Other vector processors and SIMDs with long vector lengths, such as the Intel AVX512, may also contain valid tunings, but are currently unverified.
This patch for vector processors(VP) will be distributed according to the original version of OpenMX. Please check the following correspondence table.
| Patch for VP ver. | Original OpenMX ver. | note |
|---|---|---|
| vec17 | 3.9.9 | |
| vec16 | 3.9.9 | deprecated |
| vec15 | 3.9.6 | |
| vec14 | 3.9.2 | |
| vec13 | 3.9.2 | |
| none | 3.9.1 | |
| none | 3.9.0 |
This section describes how to install the original OpenMX and how to apply this patch.
- Download the original OpenMX, the original patch, and this patch for VP. And, put them in your directory.
$ ls
openmx3.9.tar.gz patch3.9.9tar.gz openmx-vector-patch-vec17for3.9.9.tar.gz
- Expand the original OpenMX
$ tar xvfz openmx3.9.tar.gz
$ cd openmx3.9
$ ls
DFT_DATA19 source work
- Apply the original patch
$ cp -rp source source399vec17
$ cd source399vec17
$ tar xvfz ../../patch3.9.9.tar.gz
- Apply the patch for VP
$ tar xvfz ../../openmx-vector-patch-vec17for3.9.9.tar.gz
$ cp -rp openmx-vector-patch-vec17for3.9.9/source3.9.9_sx/* ./
$ rm -r openmx-vector-patch-vec17for3.9.9/
- Check and modify the makefile according to your environment. And, Make it.
$ vi makefile
$ make -j install
- Check culculation. However, follow the rules of your computer environment when submitting jobs.
$ cd ../work
# Before benchmark, check your computer rule for job submission.
$ ./openmx -runtest
$ ./openmx -runtestL
The result of the runtest, which is "runtest.result", may shows not small differences of forces for cases of CO and H2O as follows:
$cat runtest.result
1 input_example/Benzene.dat Elapsed time(s)= 13.53 diff Utot= 0.000000000005 diff Force= 0.000000000009
2 input_example/C60.dat Elapsed time(s)= 32.27 diff Utot= 0.000000000021 diff Force= 0.000000000002
3 input_example/CO.dat Elapsed time(s)= 38.20 diff Utot= 0.000000000010 diff Force= 0.002236432722
4 input_example/Cr2.dat Elapsed time(s)= 18.72 diff Utot= 0.000000000434 diff Force= 0.000000000137
5 input_example/Crys-MnO.dat Elapsed time(s)= 29.50 diff Utot= 0.000000000001 diff Force= 0.000000000034
6 input_example/GaAs.dat Elapsed time(s)= 54.21 diff Utot= 0.000000000013 diff Force= 0.000000000000
7 input_example/Glycine.dat Elapsed time(s)= 16.14 diff Utot= 0.000000000002 diff Force= 0.000000000003
8 input_example/Graphite4.dat Elapsed time(s)= 10.49 diff Utot= 0.000000000020 diff Force= 0.000000000013
9 input_example/H2O-EF.dat Elapsed time(s)= 13.29 diff Utot= 0.000000000002 diff Force= 0.000000001718
10 input_example/H2O.dat Elapsed time(s)= 11.27 diff Utot= 0.000000000002 diff Force= 0.000741132742
11 input_example/HMn.dat Elapsed time(s)= 27.99 diff Utot= 0.000000000457 diff Force= 0.000000000002
12 input_example/Methane.dat Elapsed time(s)= 10.69 diff Utot= 0.000000000004 diff Force= 0.000000000000
13 input_example/Mol_MnO.dat Elapsed time(s)= 20.34 diff Utot= 0.000000000099 diff Force= 0.000000000013
14 input_example/Ndia2.dat Elapsed time(s)= 11.34 diff Utot= 0.000000000003 diff Force= 0.000000000001
Total elapsed time (s) 307.99
This is caused by the change of algorism in "MD_pac.c" to accelarate the structure relaxation which is "MD.type == opt". These differences of forces are not error.
The contents of tuning are summarized in the following paper. When using this code, it would be appreciated if you can cite this paper.
Both the original OpenMX code and the patch for VP are licensed as GPL3.