Compiling from Source
This guide is intended to help developers to get our programs running.
Note
To install our programs refer to Setup and Installation. If you come here from Setup and Installation, since our instructions have not worked for you, why not drop us a hint at the issue tracker?
The xtb source code is hosted at GitHub.
Building with meson
The xtb program source comes with a meson build-system
(see mesonbuild.com for details).
Despite being a rather young build-system, we decided to commit to the
idea of using it for xtb due to its simplicity and speed compared
to competing build-systems like Scons or Make.
To build xtb from the source the meson build system can be used.
For a decent Fortran support verson 0.51 of meson or newer is required,
additionally the default backend ninja is required with version 1.7 or newer.
Getting meson
To install the meson build system first check your package manager for an up-to-date meson version,
usually this will also install ninja as dependency.
Alternatively you can install the latest version of meson and ninja with pip (or pip3 depending on your system):
pip install meson ninja [--user]
If you prefer conda as a package manage you can install meson and ninja from the conda-forge channel.
Make sure to select the conda-forge channel for searching packages.
conda config --add channels conda-forge
conda install meson ninja
Configure Intel Fortran build with MKL
The recommended build for xtb is with Intel Parallel Studio using the Intel Fortran compiler and the Math Kernel Library as default backend.
Precompiled, statically linked xtb binaries for Linux are provided at the release page.
The setup for the linear algebra backend defaults to MKL, therefore, only the compilers have to exported before configuring the build:
export FC=ifort CC=icc
meson setup build --buildtype=release
After the configuration step the build can be performed with ninja:
ninja -C build
Note, ninja will by default use all the threads available on your system.
Note
If you share the build machine with others it might be helpful to reduce the number of concurrent jobs using the -j flag.
By default the binary will be linked statically, other supported backends are:
backend |
linked against |
|---|---|
mkl-static |
static MKL (default) |
mkl |
shared MKL |
mkl-rt |
MKL real time library |
openblas |
OpenBLAS and if required LAPACK |
netlib |
BLAS and LAPACK |
custom |
|
Note
If you are using the MKL provided by conda-forge you have to link against the netlib backend
Configure GCC build with OpenBLAS
xtb can also be compiled with GCC version 8 or later.
For this example we additonally choose to change the linear algebra backend to OpenBLAS, if you have Intel Parallel Studio installed, you can leave out the last argument to get the MKL backend.
export FC=gfortran CC=gcc
meson setup build --buildtype=release -Dlapack=openblas
The build system will check if the OpenBLAS library provides LAPACK features as well, if this is not the case it will additionally search for LAPACK.
If you are compiling xtb on Darwin platforms, ensure that GCC is the actual GCC and not clang.
The build can be performed just like before:
ninja -C build
Testing the build with meson
After successfully building the xtb program ensure that it is working as expected. Run the testsuite with
ninja -C build test
All tests should pass, otherwise open an issue.
Installing with meson
To use xtb in production or to pack a release with precompiled binaries the project should be installed with ninja.
The installation prefix defaults to /usr/local on Linux systems, you might want to adjust this first by configuring your build with
meson configure build --prefix=$HOME/.local
To perform the actual installation run
ninja -C build install
Depending on the installation prefix and your user rights ninja might ask for the root access to perform the installation.
Building with GPU support
This projects can run on accelerator devices from NVIDIA. The compilation of the GPU version requires the NVIDIA HPC SDK, dupped NVHPC for brevity. The NVHPC compilers are available for free here.
Note
It is highly recommended to carefully compare the performance of the CPU version with the GPU version before starting production runs. Certain problem sizes can profit more from different accelerator devices than others.
To throw in some numbers as guidance for a single point calculation of a 3000 atom system with GFN2-xTB(ALPB) using xtb version 6.4.0:
Compiler |
Hardware |
Walltime |
|---|---|---|
Intel 18 |
4 cores @ Intel Xeon CPU E3-1270 v5 |
13 min |
Intel 18 |
8 cores @ Intel Xeon Gold 6148 CPU |
7 min |
NVHPC 20.7 |
Tesla K80 (cc35) |
7 min |
NVHPC 20.7 |
Tesla V100 (cc70) |
2 min |
The NVHPC provides TCL environment modules which are the preferred way to setup the compilers, if your module environment is already configured, you can just go ahead and
module load nvhpc
Note
The TCL environment modules are usually installed in the highest level of
your chosen install prefix, i.e. /opt/nvhpc/modulefiles if you
installed into /opt/nvhpc.
If you do not have a module environment available on your (local) system
you can install the TCL environment modules under Ubuntu with the
environment-module package or the newer Lua environment modules with
the lmod package.
With the NVHPC compilers available, configure a build with
export FC=nvfortran CC=nvc
meson setup build_gpu --prefix=$HOME/.local -Dlapack=netlib -Dgpu=true -Dcusolver=true
You can select the correct compute capability of your device with -Dgpu_arch=70.
Note
Support for NVHPC in meson is available since version 0.56.0.
Compile and install the project with
ninja -C build_gpu install
If you used the provided TCL environment modules of the NVHPC, you can use xtb in a similar way by including the automatically generated TCL environment module in the install prefix with:
echo "prereq nvhpc" >> ~/.local/share/modules/modulefiles/xtb/*
module use ~/.local/share/modules/modulefiles
module load xtb
Now you have a working version of xtb which can make use of your GPU.
To check if your GPU is utilized correctly you can either track the GPU usage with
nvidia-smi command line tool or set PGI_ACC_NOTIFY=3 when running xtb
as environment variable to get information on which kernels are launched on which device.
If you have multiple accelerator devices attached to your system you can select them
at runtime with CUDA_VISIBLE_DEVICES=<int>.
Supported Compilers
This is a non-comprehensive list of tested compilers for xtb with the meson build system.
Compiler |
Version |
Platform |
Architecture |
|
|---|---|---|---|---|
GCC |
10.2 |
Ubuntu 20.04 |
x86_64 |
6.4.1, latest |
GCC |
10.2, 11.1 |
Manjaro Linux |
x86_64 |
6.4.0, 6.4.1 |
GCC |
10.2 |
Windows Server 2019 |
x86_64 |
6.4.0, 6.4.1, latest |
GCC |
9.3 |
Ubuntu 18.04 |
x86_64 |
6.4.0 |
GCC |
9.3 |
Ubuntu 20.04 |
x86_64 |
6.4.1, latest |
GCC |
9.3 |
Centos 7 |
ppc64le, aarch64 |
6.4.0, 6.4.1 |
GCC |
9.3 |
Centos 6 |
x86_64 |
6.4.0, 6.4.1 |
GCC |
9.3 |
MacOS 10.15.7 |
x86_64 |
6.4.0, 6.4.1, latest |
GCC |
8.4 |
Ubuntu 20.04 |
x86_64 |
6.4.1, latest |
GCC |
7.5 |
Ubuntu 18.04 |
x86_64 |
6.4.0 |
Intel |
2021.1, 2021.2 |
Ubuntu 20.04 |
x86_64 |
6.4.0, 6.4.1, latest |
Intel |
18.0.2 |
OpenSuse 42.1 |
x86_64 |
6.4.0, 6.4.1 |
NVHPC |
20.11, 21.1 |
Manjaro Linux |
x86_64 |
6.4.0 |
NVHPC |
20.9 |
Centos 8 |
x86_64 + cc70 |
6.4.0 |
NVHPC |
20.7 |
OpenSuse 42.1 |
x86_64 + cc35 |
6.4.0 |
The list was started with version 6.4.0 and will be continued for future released.
The latest version refers to the continuously tested compiler tool chains in the xtb repository.
For GPU enabled builds the compute-capability is given together with the architecture.
Note
First class compiler support in xtb comes only with continuous testing, if you
want to see a particular compiler, platform or architecture in the list above,
please reach out to us at the discussion board,
open an issue or submit
a continuous integration workflow with a pull request to xtb.