Caution
This branch is under heavy development. Many features are missing or not fully implemented. If you are not an ARTS developer, you might want to stay away.
ARTS is free software. Please see the file COPYING for details.
If you use data generated by ARTS in a scientific publication, then please mention this and cite the most appropriate of the ARTS publications that are summarized on http://www.radiativetransfer.org/docs/
CONTRIBUTING.md provides information on contributing to ARTS on GitHub.
For documentation, please see the files in the doc subdirectory.
For building and installation instructions please read below.
Build Prerequisites (provided by Miniforge3):
See the platform-dependent environment-dev-{linux,mac,win}.yml file for all requirements.
The following instructions assume that you are using Miniforge3 as a build environment. The installer is available at the project's Github page.
Use the provided environment-dev-{linux,mac,win}.yml files to install
all required dependencies into your current conda environment.
Optionally, a separate environment for development can be created, if you want to keep your current environment clean:
mamba create -n pyarts-dev
mamba activate pyarts-dev
Install dependencies on Linux:
mamba env update -f environment-dev-linux.yml
Install dependencies on macOS:
mamba env update -f environment-dev-mac.yml
Install dependencies on Windows:
mamba env update -f environment-dev-win.yml
Next, follow these steps to use cmake to build ARTS:
cd arts
cmake --preset=default-gcc-conda # On macOS use default-clang-conda, on windows use default-msvc-conda
cmake --build build -jX
X is the number of parallel build processes. X=Number of Cores gives you usually the fastest compilation time.
WARNING: The compilation is very memory intensive. If you have 16GB of RAM, don't use more than 6-8 cores. With 8GB, don't use more than 2-3 cores.
Development install of the PyARTS Python package:
python3 -m pip install --user -e build/python
You only have to do the python package install once. If the ARTS source has changed, update the PyARTS package by running:
cmake --build build -jX --target pyarts
By default, ARTS is built in release mode with optimizations enabled and assertions and debugging symbols turned off.
Whenever you change the configuration, remove your build directory first:
rm -rf build
To build with assertions and debugging symbols use:
cmake --preset=default-gcc-conda -DCMAKE_BUILD_TYPE=RelWithDebInfo
This configuration offers a good balance between performance and debugging capabilities. Since this still optimizes out many variables, it can be necessary for some debugging cases to turn off all optimizations. For those cases, the full debug configuration can be enabled. Note that ARTS runs a lot slower in this configuration:
cmake --preset=default-gcc-conda -DCMAKE_BUILD_TYPE=Debug
To install the PyARTS Python package, you need to build it and install it with pip. Create your build directory and configure ARTS with cmake as described in the previous sections. Then, run the following commands inside your build directory:
cmake --build build --target pyarts
python3 -m pip install --user -e build/python
This will not mess with your system's Python installation.
A link to the pyarts package is created in your home directory, usually
$HOME/.local/lib/python3.X/site-packages/pyarts.egg-link.
You don't need to reinstall the package with pip after updating ARTS.
You only need to run cmake --build build --target pyarts again.
'cmake --build build --target check' will run several test cases to ensure that ARTS is working properly. Use 'check-all' to run all available controlfiles, including computation time-intensive ones.
Some tests depend on the arts-xml-data package. cmake automatically looks if it is available in the same location as ARTS itself. If necessary, a custom path can be specified.
cmake --preset=default-gcc-conda -DARTS_XML_DATA_PATH=/home/myname/arts-xml-data
If arts-xml-data cannot be found, those tests are ignored.
By default, 4 tests are executed in parallel.
If you change the number of concurrently run test, you can add this option to your cmake --preset=.... call:
-DTEST_JOBS=X
X is the number of tests that should be started in parallel.
You can also use the ctest command directly to run the tests:
First, change to the build directory:
cd build
This runs all test with 4 jobs concurrently:
ctest -j4
To run specific tests, use the -R option and specify part of the test case name you want to run. The following command will run all tests that have 'ppath' in their name, e.g. arts.ctlfile.fast.ppath1d ...:
ctest -R ppath
To see the output of ARTS, use the -V option:
ctest -V -R fast.doit
By default, ctest will not print any output from ARTS to the screen. The option --output-on-failure can be passed to ctest to see output in the case an error occurs. If you want to always enable this, you can set the environment variable CTEST_OUTPUT_ON_FAILURE:
export CTEST_OUTPUT_ON_FAILURE=1
To squeeze out every last drop of performance, you can also build a version specifically optimized for your machine's processor:
-DCMAKE_BUILD_TYPE=Native
This option should make the executable slightly faster, more so on better systems, but not portable. Note that since this build-mode is meant for fast-but-accurate computations, some IEEE rules will be ignored. For now only complex computations are IEEE incompatible running this mode of build.
Enable NetCDF: The basic matpack types can be read from NetCDF files, if NetCDF support is enabled:
cmake --preset=default-gcc-conda -DENABLE_NETCDF=1
Precompiled headers: PCH can speed up builds significantly. However, it hampers the ability for ccache to properly skip unnecessary compilations, potentially increasing rebuild times. Tests have shown that it only speeds up the build considerably for Clang, but not for GCC.
cmake --preset=default-clang-conda -DENABLE_PCH=1 ..
If you enable PCH and also use ccache, you need to set the CCACHE_SLOPPINESS
environment variable properly:
export CCACHE_SLOPPINESS=pch_defines,time_macros
Disable assertions: -DNO_ASSERT=1
Disable OpenMP: -DNO_OPENMP=1
Disable the built-in documentation server: -DNO_DOCSERVER=1
The build utilizes ccache automatically when available, it can be
turned off with the option -DENABLE_CCACHE=0
For details see https://ccache.samba.org/
The callgrind plugin included in valgrind is the recommended profiling method for ARTS.
Certain things should be taken into account when calling ARTS with valgrind. Since recursion (cycles) will lead to wrong profiling results it is important to use the following settings to obtain profile data for ARTS:
valgrind --tool=callgrind --separate-callers=10 --separate-recs=3 arts -n1 ...
For detail on these options consult the valgrind manual:
http://valgrind.org/docs/manual/cl-manual.html#cl-manual.cycles
-n1 should be passed to ARTS because parallelisation can further scew the results. Since executing a program in valgrind can lead to 50x slower execution, it is recommended to create a dedicated, minimal controlfile for profiling.
After execution with valgrind, the resulting callgrind.out.* file can be opened in kcachegrind[1] for visualization. It is available as a package for most Linux distributions.
Note that you don't have to do a full ARTS run. You can cancel the program after some time when you think you have gathered enough statistics.
[1] https://kcachegrind.github.io/
The Performance Counters for Linux offer a convenient way to profile any program with basically no runtime overhead. Profiling works for all configurations (Debug, RelWithDebInfo and Release). To ensure that the calltree can be analyzed correctly, compile ARTS with frame pointers. This has minimal impact on performance. Use the following preset to enable this setting:
cmake --preset=perf-gcc-conda
Prepend the perf command to your arts call to record callgraph information:
perf record -g src/arts MYCONTROLFILE.arts
This can also be applied to any test case:
perf record -g ctest -R TestDOIT$
After recording, use the report command to display an interactive view of the profiling information:
perf report -g graph,0.5,callees
This will show a reverse call tree with the percentage of time spent in each function. The function tree can be expanded to expose the calling functions.