15.4. Cmake for AFNI - making AFNI is a piece of C(m)ake¶
Notes by J Lee.
In addition to the build systems using make files, you can build the AFNI project using the CMake build system. This build system offers several advantages some of which are:
Parallel builds can be performed greatly accelerating build time (at least an order of magnitude if you have the CPUs)
One build to rule them all… it configures itself for Linux or Mac systems so you don’t have to tweak it for your specific system. This works for Clang, icc, gcc (the latter most frequently used)
Dependency management: as part of the previous point, you can detect and build against dependencies on your system. Alternatively you can specify CMake options to directly download, build, and install dependencies from source
Shared object linking: CMake manages shared object linking (linking against .so or .dylib files) in a robust way. Depending on whether you have a build directory or have installed the software, the rpath of all binaries is set appropriately to link against all dependencies.
# change to your home directory, or whatever you would prefer cd git clone https://github.com/afni/afni cd afni mkdir build; cd build cmake [options] .. # can make specific targets like afni,3dDeconvolve,etc make
All binaries should now be present in
./targets_built (in build
directory). Note that scripts (tcsh, R, python etc.) will remain in
the source tree until installation.
N.B. For a faster build you can instead use the ninja build system set the environment variable CMAKE_GENERATOR to “Ninja” and substitute the make command above for “ninja” (cmake v>3.14). But see Section: Ninja build system.
For a more comprehensive overview see the Section: Basic approach to building and installing.
In order to get a basic working setup on your computer you can try the suggestions below :
apt-get install ca-certificates curl freeglut3-dev libf2c2-dev \ libglib2.0-dev libglu1-mesa-dev libglw1-mesa-dev \ libgsl-dev libgts-dev libjpeg62 libmotif-dev libnetcdf-dev \ libxi-dev libxmhtml-dev libxmu-dev libxpm-dev libxt-dev \ libvolpack1-dev python-dev python3.6-dev qhull-bi \ r-base tcsh ninja-build
GLW breaks our build on ubuntu/debian. Use neurodebian or use the
brew cask install xquartz brew unlink python\@2 brew install \ llvm cmake ninja pkgconfig \ jpeg gsl gts openmotif netcdf libpng expat \ freetype fontconfig gsl netpbm git-annex
Note, for a more comprehensive list of development dependencies it may be worth checking out the files used for continuous integration testing: github.com/afni/.circleci/config.yml and github.com/afni/.docker/afni_dev_base.dockerfile
Cmake detects all of the details of your system and then generate a build system (for example it will write Make files) that can subsequently be executed.
When running cmake, it will try to find out the details of your system based on the options you have passed on the command line. This includes checking that required compiler properties and other dependencies are acceptable on the system. The most common errors at this point would be that you do not have a dependency installed.
Assuming no errors occur during cmake’s configure-time, cmake will try to generate a build system. This consists of writing a set of build files (for example Make files) to the build directory that will execute on the current host.
After the build system is generated you can build your project in the conventional way. I.e. for a Make build system you simply execute:
This will generate all of the binaries and place them in the targets_built subdirectory of the build directory (removing any of these binaries will trigger a rebuild for these specific binaries and any of their dependents). You can run the executables as expected by simply typing something like
You can use this by setting the environment variable CMAKE_GENERATOR
to “Ninja” (cmake version > 3.14) or by adding
-GNinja to the
cmake command. There are many performance advantages to using ninja,
the most notable being that a
no op build is close to
instantaneous for “ninja” whereas for “make” the equivalent state
takes approximately a minute to determine that nothing is currently
required of the build system.
One potential issue is that ninja automatically computes the number of threads it “should” use in parallel. On some Macs this seems to cause them to crash. This may be a memory issue. It could potentially be resolved by using ninja’s job pool functionality. For now the issue can be fixed using the -j flag to tell ninja to set the number of threads used for the build. The optimal number of threads to use can be figured out through experimentation.
WARNING: By default an installation will go into /usr/local which you
likely do not want to do. When testing that the installation works this
installation location can be overwritten easily by setting the
For example in bash:
DESTDIR=local_install_dir make install
setenv DESTDIR local_install_dir make install
It is worth noting that a build target (pytest) exists to do inplace testing on the build. As one might expect it uses the pytest software to run these tests. This setup may obviate any need to do an install as part of your development workflow. This has the advantage that it always checks that the project is up to date before running any tests, it manages test data, and it temporarily modifies the PATH in order to have all AFNI executables available for testing (both built binaries and scripts in the source tree). For further details have a look at the Section: Running tests.
The bin subdirectory of an installation should be added to one’s path to make use of the “installed” AFNI. Note that this does not necessarily have to be installed into system directories. Once on the path, you should have access to all of the executables expected from a full AFNI installation (as in tcsh, R… with a cmake option, and python executables are not available following a build but they are available after an install). If you observe any behavior that deviates from a standard AFNI install please raise an issue on github.
This section is for when you have added a new software tool and you wish to incorporate it into the cmake build.
In brief if I want to add a new executable, my_new_binary, using my new
source code in
src/new_exec.c then I would add the cmake code:
Executables, libraries, or plugins are added by using the cmake
add_afni_plugin, respectively. These are wrappers around the
standard equivalents of cmake
As with the make build system, CMake refers to all of these entities
that require compilation as “targets”.
The cmake files for adding targets all follow the pattern
CMake*.txt. It is usually fairly obvious which CMake file you
should use for your new program (this is purely convention). For
example, in general for adding new binary programs use the
CMake_binaries.txt file. This will help to keep things more structured
and typically there are many examples in the appropriate files to help
you deal with tricky details. An attempt at an index of such details
is in Section: Examples.
Once you have correctly added a new target you will have to consider updating the list of expected targets (see Section: List of expected targets)
The most common error will be missing dependencies. I have currently attempted to mitigate this by setting the defaults to just work. Failures in this should be reported. In attempting to resolve this yourself you can attempt the following.
Try to install the missing software. Hopefully, the missing package will be fairly self-explanatory from the error message. The base dockerfile should give you an idea of the dependencies that need to be satisfied in order to fully build and test AFNI.
You can try to use a build of the dependency from the AFNI repo/cmake driven source code download. At the end of the cmake options file (github.com/afni/cmake/afni_cmake_build_options.cmake) you can see many packages that can be installed using this alternative strategy. The basic approach is to add a flag to the cmake command to avoid trying to find the system installed version of the software...
There are situations in which the dependency resolution can be a lot more difficult. Getting help in those situations is probably best. The main issue would be that some software is in fact installed on the system but it is not detected. This is would be a bug in the cmake system and should be fixed.
The cmake system has “targets” for the various programs and libraries. The cmake build is set up to attempt build all targets to achieve feature parity with the Make build. There are fairly aggressive safeguards to try to enforce synchronization of the two builds. If targets are built that are not expected, or expected targets are not built, the cmake system will raise an error at configure time. This can be frustrating but will hopefully be relaxed in the future when a full transition to the cmake build has occurred.
There are three ways of keeping track of the targets in the AFNI project:
Extracting a list of targets build by the pre-existing Make build system
Checking the contents of packaging/installed_components.txt
Extracting a list of targets built by the cmake generated build system
During the cmake build the contents of the category 3 is compared with that of category 2 and an error occurs if the two lists do not match. 3 is compared to 1, and a report of the differences are detailed in the cmake output to help determine divergence in the two build systems.
Warning: The details of this section are encapsulated in the run_afni_tests.py script in the tests directory. You may not wish to read this.
A more extensive/up-to-date description can be found at this link.
A build target (pytest) exists to do inplace testing on the build. This target uses the pytest software to run these tests. This setup may obviate any need to do an install as part of your development workflow. This has the advantage that it always checks that the project is up to date before running any tests, it manages test data, and it temporarily modifies the PATH in order to have all AFNI executables available for testing (both built binaries and scripts in the source tree). For further details have a look at the running tests section
The ARGS environment variable can be set to modify the behavior of this target. Examples:
export ARGS='scripts --workers 3 -k mask' ninja pytest
The bin subdirectory of a build should be added to one’s path to make use of the “installed” AFNI. Note this may not be installed into system directories. This will give access to all of the executables expected from a full AFNI installation (as in tcsh, R, and python executables are installed into bin but they do not get copied into the build output directory).
The basic system setup on neurodebian for both make and cmake builds can be seen here (it is the instructions used to build the base docker image for both builds): github.com/afni/.docker/afni_dev_base.dockerfile
The cmake build on neurodebian can be seen in the cmake dockerfile: github.com/afni/afni/.docker/cmake_build.dockerfile
A build on MacOS occurs on CircleCI: github.com/afni/.circleci/config.yml
The books Professional CMake and CMake Cookbook are both excellent. The former serves as an in-depth advanced reference. The latter has many useful examples that are carefully explained.
Testing documentation is at this link.
Sorted somewhat by order of frequency it is required:
Adding targets whose .h files do not match the name of the
Installing scripts and other files:
Specifying settings with default values that can be used throughout the build:
Linking against external libraries:
Setting compile time definitions for targets:
Setting compile time definitions for specific source files:
Specifying public and private dependencies of libraries to conveniently propagate compilation/link settings to dependent libraries/executables
Adding headers that are needed at compilation but shouldn’t be distributed elsewhere:
Using INTERFACE libraries to establish compile definitions, headers etc but doesn’t actually get created by the build system:
Writing custom commands for configure time
Writing custom commands for build time
Creating “object” libraries (a collection of .o files for convenience that can be reused across binaries)
Dealing with targets whose source files span several directories
Specifying linking in a conditional way dependending on system, build configuration:
*.cfiles as similar to header files in that they are included and dependent targets should also be able to include them:
Using external libraries that have diverged slightly from versions now distributed by package-managers:
Encapsulating code in functions, despite the weird scoping rules of the cmake language:
Running build time checks on compiled binaries: TODO, would use add_custom_command:
Running build time check for missing symbols
Building with OMP support:
Modifying the toolchain to deal with switching compilers: