tobigithub
commented
4 years ago One can also use netlib lapack backend, the best remove the whole build directory and do a fresh unzip. The reconfigure command for meson and ninja also works.
>meson --reconfigure build_gcc/
>ninja -C build_gcc testor
>unzip master.zip
>cd xtb-master
>export FC=gfortran-8 CC=gcc-8
>meson setup build_gcc --buildtype release -Dla_backend=netlib --warnlevel 0
>ninja -C build_gcc testwe get the same results that all tests were OK
Ok: 54
Expected Fail: 5
Fail: 0
Unexpected Pass: 0
Skipped: 3
Timeout: 0
Full log written to /home/tkind/xtb_source/xtb-master/build_gcc/meson-logs/testlog.txtthen we can test xtb with the caffeine example:
tkind@instance-1:~/xtb_source/test/caffeine-netlib$ ./xtb --version
-----------------------------------------------------------
| ===================== |
| x T B |
| ===================== |
| S. Grimme |
| Mulliken Center for Theoretical Chemistry |
| University of Bonn |
-----------------------------------------------------------
* xtb version 6.2.3 (unknown) compiled by 'tkind@instance-1' on 2020-03-20
normal termination of xtb
>./xtb --opt extreme caffeine.coord
optimized geometry written to: xtbopt.coord
-------------------------------------------------
| TOTAL ENERGY -42.153937410494 Eh |
| GRADIENT NORM 0.000011585590 Eh/α |
| HOMO-LUMO GAP 3.424625890456 eV |
-------------------------------------------------
------------------------------------------------------------------------
* finished run on 2020/03/20 at 04:21:22.044
------------------------------------------------------------------------
total:
* wall-time: 0 d, 0 h, 0 min, 0.816 sec
* cpu-time: 0 d, 0 h, 0 min, 0.761 sec
* ratio c/w: 0.932 speedup
SCF:
* wall-time: 0 d, 0 h, 0 min, 0.050 sec
* cpu-time: 0 d, 0 h, 0 min, 0.044 sec
* ratio c/w: 0.891 speedup
ANC optimizer:
* wall-time: 0 d, 0 h, 0 min, 0.727 sec
* cpu-time: 0 d, 0 h, 0 min, 0.690 sec
* ratio c/w: 0.950 speedup
normal termination of xtb
Note: The following floating-point exceptions are signalling: IEEE_UNDERFLOW_FLAG IEEE_DENORMALWhile the file size is the same (netlib/openblas) the md5sum is different. Also the executables seem small compared to the intel/mkl compiled binaries, so maybe they are not statically linked.
compilation of gcc on ubuntu (VM instance on GCP)
Compilation can be done with meson/ninja, or with make/cmake. Here is the complete installation procedure from a blank system.
The refer to build instructions: https://xtb-docs.readthedocs.io/en/latest/development.html
1) Install pip3, ninja and meson, be aware meson and ninja in this case are installed via pip3 not natively (apt install meson) on the OS level. In the case below, without the sudo command meson is installed into "~/.local/bin/meson". See https://github.com/mesonbuild/meson/issues/1613 and https://mesonbuild.com/Quick-guide.html
2) Install the BLAS libraries under Ubuntu, if openblas is not installed it may raise an error "ERROR: Fortran library 'openblas' not found"
3) Install gcc and gfortran (latest >8.x) see https://www.osradar.com/install-gnu-fortran-on-ubuntu-18-04/ the normal "sudo apt-get install gfortran" only installs gfortran 7.5 hence we need gfortran-8 this can also be confirmed by looking into the xtb repository
4) get the xtb sources, with wget on gets the master.zip (Windows gets xtb-master)
5) make changes to the system and then built (UBUNTU gcc), openblas backend is defined here, example for netlib see below. https://github.com/grimme-lab/xtb/blob/master/meson_options.txt also see https://github.com/mesonbuild/meson/issues/4890
6) After running ninja with the tests, a number of warnings may occur, not sure if relevant or not or covered by the tests. Compile time on one Intel Xeon core (2.4 Ghz) is around 1-2 minutes.
7) Then copy the xtb binary from the folder "/xtb_source/xtb-master/build_gcc/" and run a version test and run a caffeine test.
8) Now testing the xtb computations whith caffeine as example:
9) running xtb (just copied into the same folder, observe the binaries and config files and folders when transferring)
Comparing with INTEL compilers and MKL
Basically the difference between the INTEL compiler and gcc is 0.00000002719 Eh which translates into 0.000017 kcal/mol or a very small error however the IEEE_UNDERFLOW_FLAG IEEE_DENORMAL is somewhat unsettling, reminding me of Intel 387 coprocessor testing times. A quick google search reveals that there are potential remedies to this warning, which I will not explore.
Verdict compiling xtb with gcc and gfortran-8 works.