Trilinos Configure, Build, Test, and Install Reference Guide

Author: Roscoe A. Bartlett (


This document contains reference information on how to configure, build, test, and install Trilinos using the TriBITS CMake build system. The primary audience are users of Trilinos that need to configure and build the software. The secondary audience are actual developers of Trilinos.


1   Introduction

Trilinos contains a large number of packages that can be enabled and there is a fairly complex dependency tree of required and optional package enables. The following sections contain fairly generic information on how to configure, build, test, and install Trilinos that addresses a wide range of issues.

This is not the first document that a user should read when trying to set up to install Trilinos. For that, see the INSTALL.* file in the base Trilinos source directory. There is a lot of information and activities mentioned in this reference that most users (and even some Trilinos developers) will never need to know about.

Also, this particular reference has no information at all on what is actually in Trilinos. For that, go to:

to get started.

2   Trilinos-specific options

Below, configure options specific to Trilinos are given. The later sections give more generic options that are the same for all TriBITS projects.

2.1   Enabling float and complex Scalar types

Many of the packages in Trilinos are implemented using C++ templates and therefore support a variety of data-types. In addition to the default scalar type double, many of the data-structures and solvers are tested with the types float, std::complex<float>, and std::complex<double>. In addition, these packages support the explicit template instantiation (i.e. -DTrilinos_EXPLICIT_TEMPLATE_INSTANTIATION=ON) of these types as well. However, support and explicit instantiations for these types are off by default since most users don't need these extra types and enabling them greatly increases the compilation time for Trilinos libraries and tests and can consume a great deal more disk space. But support for these types in the various Trilinos packages can be enabled using the following options:


Enables suppport and explicit instantiations for the float scalar data-type in all supported Trilinos packages.


Enables suppport and explicit instantiations for the std::complex<T> scalar data-type in all supported Trilinos packages.


Enables suppport and explicit instantiations for the std::complex<float> scalar data-type in all supported Trilinos packages. This is set to ON by default when -DTrilinos_ENABLE_FLOAT=ON and -DTrilinos_ENABLE_COMPLEX=ON are set.


Enables suppport and explicit instantiations for the std::complex<double> scalar data-type in all supported Trilinos packages. This is set to ON by default when -DTrilinos_ENABLE_COMPLEX=ON is set.

2.2   Enabling/disabling time monitors

I order to enable instrumentation of select code to generate timing statistics, set:


This will enable Teuchos time monitors by default in all Trilinos packages that support them. To print the timers at the end of the program, call Teuchos::TimeMonitor::summarize().

In order do co-development of TriBTS and Trilinos (see http://, set:


(NOTE: You have to use the data-type STRING with Trilinos_TRIBITS_DIR or CMake will automatically assume it is relative to the build dir!)

3   Getting set up to use CMake

Before one can configure Trilinos to be built, one must first obtain a version of CMake on the system newer than 2.8.11 This guide assumes that once CMake is installed that it will be in the default path with the name cmake.

3.1   Installing a binary release of CMake [casual users]

Download and install the binary (version 2.8.11 or greater is recommended) from:

3.2   Installing CMake from source [developers and experienced users]

If you have access to the Trilinos git repositories, then install CMake with:

$ $TRIBITS_BASE_DIR/python/ \
   --install-dir=<INSTALL_BASE_DIR> \

This will result in cmake and related CMake tools being installed in <INSTALL_BASE_DIR>/bin.

Getting help for installing CMake with this script:

$ $TRIBITS_BASE_DIR/python/ --help

NOTE: you will want to read the help message about how to use sudo to install in a privileged location (like the default /usr/local/bin).

4   Getting CMake Help

4.2   Building CMake help locally

To get help on CMake input options, run:

$ cmake --help

To get help on a single CMake function, run:

$ cmake --help-command <command>

To generate the entire documentation at once, run:

$ cmake --help-full

(Open your web browser to the file

5   Configuring (Makefile Generator)

While CMake supports a number of different build generators (e.g. Eclipes, XCode, MS Visual Studio, etc.) the primary generator most people use on Unix/Linix system is make and CMake generates exceptional Makefiles. The materila in this section, while not exclusing to the makefile generator this should be assumed as the default.

5.1   Setting up a build directory

In order to configure, one must set up a build directory. Trilinos does not support in-source builds so the build tree must be seprate from the source tree. The build tree can be created under the source tree such as with:

$ mkdir <SOME_BUILD_DIR>

but it is generally recommended to create a build directory parallel from the source tree.

NOTE: If you mistakenly try to configure for an in-source build (e.g. with 'cmake .') you will get an error message and instructions on how to resolve the problem by deleting the generated CMakeCache.txt file (and other generated files) and then follow directions on how to create a different build directory as shown above.

5.2   Basic configuration

  1. Create a 'do-configure' script such as [Recommended]:

    cmake \
      -D Trilinos_ENABLE_TESTS=ON \
      $EXTRA_ARGS \

and then run it with:


where <TRIBITS_PACKAGE> is a valid SE Package name (see above), etc. and SOURCE_BASE is set to the Trilinos source base directory (or your can just give it explicitly in the script).

See Trilinos/sampleScripts/*cmake for examples of real do-configure scripts for different platforms..

NOTE: If one has already configured once and one needs to configure from scratch (needs to wipe clean defaults for cache variables, updates compilers, other types of changes) then one will want to delete the local CMakeCache.txt and other CMake-generated files before configuring again (see Reconfiguring completely from scratch).

  1. Create a CMake file fragment and point to it [Recommended].

Create a do-configure script like:


cmake \
  -D Trilinos_CONFIGURE_OPTIONS_FILE:FILEPATH=MyConfigureOptions.cmake \
  -D Trilinos_ENABLE_TESTS=ON \

where MyConfigureOptions.cmake (in the current working directory) might look like:

SET(CMAKE_BUILD_TYPE DEBUG CACHE STRING "Set in MyConfigureOptions.cmake")
SET(Trilinos_ENABLE_CHECKED_STL ON CACHE BOOL "Set in MyConfigureOptions.cmake")
SET(BUILD_SHARED_LIBS ON CACHE BOOL "Set in MyConfigureOptions.cmake")

Using a configuration fragment file allows for better reuse of configure options across different configure scripts and better version control of configure options. Using the comment "Set in MyConfigureOptions.cmake" makes it easy see where that variable got set when looking an the CMakeCache.txt file. Also, when this file changes, CMake will automatically trigger a reconfgure during a make (because it knows about the file and will check its time stamp).

One can use the FORCE option in the SET() shown above and that will override any value of the options that might already be set. However, that will not allow the user to override the options on the CMake comamndline using -D<VAR>=<value> so it is generally not desired to use FORCE.

One can actually pass in a list of configuration fragment files separated by ''","'' which will be read in the order they are given:

-D Trilinos_CONFIGURE_OPTIONS_FILE=<file0>,<file1>,...

One can read in configure option files under the project source directory by using the type STRING such as with:

-D Trilinos_CONFIGURE_OPTIONS_FILE:STRING=cmake/MpiConfig1.cmake

In this case, the relative paths will be with respect to the project base source directory, not the current working directroy. (By specifying the type STRING, one turns off CMake interpretation as a FILEPATH. Otherwise, the type FILEPATH causes CMake to always interpret relative paths with respect to the current working directory and set the absolute path).

Note that a CMake options file can also be read in with:

cmake -C MyConfigureOptions.cmake [other options]  ${SOURCE_BASE}

However, the advantages of using Trilinos_CONFIGURE_OPTIONS_FILE are:

1) One can use -DTrilinos_CONFIGURE_OPTIONS_FILE:STRING=<rel-path> to use a relative path w.r.t. to the source tree to make it easier to point to options files in the project source. Using cmake -C would require having to give the absolute path or a longer relative path from the build directory back to the source directory.

2) Trilinos_CONFIGURE_OPTIONS_FILE accepts a list of files where cmake -C only accepts a single file. That saves from having to create another dummy *.cmake file that just includes the others.

3) One can create and use parameterized *.cmake files that can be used with multiple TriBITS projects. For example, one can have set statements like SET(${PROJECT_NAME}_ENABLE_Fortran OFF ...) since PROJECT_NAME is known before the file is included. One can't do that with cmake -C and instead would have to the full variables names.

  1. Using the QT CMake configuration GUI:
On systems where the QT CMake GUI is installed (e.g. Windows) the CMake GUI can be a nice way to configure Trilinos (or just explore options) if you are a user. To make your configuration easily repeatable, you might want to create a fragment file and just load it by setting Trilinos_CONFIGURE_OPTIONS_FILE (see above) in the GUI.

5.3   Selecting the list of packages to enable

The Trilinos project is broken up into a set of packages that can be enabled (or disabled). For details and generic examples, see Package Dependencies and Enable/Disable Logic and TriBITS Dependency Handling Behaviors.

See the following use cases:

5.3.1   Determine the list of packages that can be enabled

In order to see the list of available Trilinos SE Packages to enable, just run a basic CMake configure, enabling nothing, and then grep the output to see what packages are available to enable. The full set of defined packages is contained the lines starting with 'Final set of enabled SE packages' and 'Final set of non-enabled SE packages'. If no SE packages are enabled by default (which is base behavior), the full list of packages will be listed on the line 'Final set of non-enabled SE packages'. Therefore, to see the full list of defined packages, run:

./do-configure 2>&1 | grep "Final set of .*enabled SE packages"

Any of the packages shown on those lines can potentially be enabled using -D Trilinos_ENABLE_<TRIBITS_PACKAGE>=ON (unless they are set to disabled for some reason, see the CMake output for package disable warnings).

Another way to see the full list of SE packages that can be enabled is to configure with Trilinos_DUMP_PACKAGE_DEPENDENCIES = ON and then grep for Trilinos_SE_PACKAGES using, for example:

./do-configure 2>&1 | grep "Trilinos_SE_PACKAGES: "

5.3.3   Enable a set of packages

To enable an SE package <TRIBITS_PACKAGE> (and optionally also its tests and examples), configure with:


This set of arguments allows a user to turn on <TRIBITS_PACKAGE> as well as all packages that <TRIBITS_PACKAGE> can use. All of the package's optional "can use" upstream dependent packages are enabled with -DTrilinos_ENABLE_ALL_OPTIONAL_PACKAGES=ON. However, -DTrilinos_ENABLE_TESTS=ON will only enable tests and examples for <TRIBITS_PACKAGE> (or any other packages specifically enabled).

If a TriBITS package <TRIBITS_PACKAGE> has subpackages (e.g. <A>, <B>, etc.), then enabling the package is equivalent to setting:


However, a TriBITS subpackage will only be enabled if it is not already disabled either explicitly or implicitly.

NOTE: The CMake cache variable type for all XXX_ENABLE_YYY variables is actually STRING and not BOOL. That is because these enable variables take on the string enum values of "ON", "OFF", end empty "". An empty enable means that the TriBITS dependency system is allowed to decide if an enable should be turned on or off based on various logic. The CMake GUI will enforce the values of "ON", "OFF", and empty "" but it will not enforce this if you set the value on the command line or in a SET() statement in an input `*.cmake options files. However, setting -DXXX_ENABLE_YYY=TRUE and -DXXX_ENABLE_YYY=FALSE is allowed and will be interpreted correctly..

5.3.4   Enable to test all effects of changing a given package(s)

To enable an SE package <TRIBITS_PACKAGE> to test it and all of its down-stream packages, configure with:


The above set of arguments will result in package <TRIBITS_PACKAGE> and all packages that depend on <TRIBITS_PACKAGE> to be enabled and have all of their tests turned on. Tests will not be enabled in packages that do not depend on <TRIBITS_PACKAGE> in this case. This speeds up and robustifies pre-push testing.

5.3.5   Enable all packages with tests and examples

To enable all SE packages (and optionally also their tests and examples), add the configure options:


Specific packages can be disabled with Trilinos_ENABLE_<TRIBITS_PACKAGE>=OFF. This will also disable all packages that depend on <TRIBITS_PACKAGE>.

All examples are also enabled by default when setting Trilinos_ENABLE_TESTS=ON.

By default, setting Trilinos_ENABLE_ALL_PACKAGES=ON only enables primary tested (PT) code. To have this also enable all secondary tested (ST) code, one must also set Trilinos_ENABLE_SECONDARY_TESTED_CODE=ON.

NOTE: If the project is a "meta-project", then Trilinos_ENABLE_ALL_PACKAGES=ON may not enable all the SE packages but only the project's primary meta-project packages. See Package Dependencies and Enable/Disable Logic and TriBITS Dependency Handling Behaviors for details.

5.3.6   Disable a package and all its dependencies

To disable an SE package and all of the packages that depend on it, add the configure options:


For example:


will enable <PACKAGE_A> and all of the packages that it depends on except for <PACKAGE_B> and all of its forward dependencies.

If a TriBITS package <TRIBITS_PACKAGE> has subpackages (e.g. <A>, <B>, etc.), then disabling the package is equivalent to setting:


The disable of the subpackage is this case will override any enables.

If a disabled package is a required dependency of some explicitly enabled downstream package, then the configure will error out if Trilinos_DISABLE_ENABLED_FORWARD_DEP_PACKAGES=OFF. Otherwise, a WARNING will be printed and the downstream package will be disabled and configuration will continue.

5.3.7   Remove all package enables in the cache

To wipe the set of pakage enables in the CMakeCache.txt file so they can be reset again from scratch, configure with:

$ ./-do-confiugre -D Trilinos_UNENABLE_ENABLED_PACKAGES=TRUE

This option will set to empty '' all package enables, leaving all other cache variables as they are. You can then reconfigure with a new set of package enables for a different set of packages. This allows you to avoid more expensive configure time checks and to preserve other cache variables that you have set and don't want to loose. For example, one would want to do this to avoid compiler and TPL checks.

5.4   Selecting compiler and linker options

The compilers for C, C++, and Fortran will be found by default by CMake if they are not otherwise specified as described below (see standard CMake documentation for how default compilers are found). The most direct way to set the compilers are to set the CMake cache variables:

-D CMAKE_<LANG>_COMPILER=<path-to-compiler>

The path to the compiler can be just a name of the compiler (e.g. -DCMAKE_C_COMPILER=gcc) or can be an absolute path (e.g. -DCMAKE_C_COMPILER=/usr/local/bin/cc). The safest and more direct approach to determine the compilers is to set the absolute paths using, for example, the cache variables:

-D CMAKE_C_COMPILER=/opt/my_install/bin/gcc \
-D CMAKE_CXX_COMPILER=/opt/my_install/bin/g++ \
-D CMAKE_Fortran_COMPILER=/opt/my_install/bin/gfortran

or if TPL_ENABLE_MPI=ON (see Configuring with MPI support) something like:

-D CMAKE_C_COMPILER=/opt/my_install/bin/mpicc \
-D CMAKE_CXX_COMPILER=/opt/my_install/bin/mpicxx \
-D CMAKE_Fortran_COMPILER=/opt/my_install/bin/mpif90

If these the CMake cache variables are not set, then CMake will use the compilers specified in the environment variables CC, CXX, and FC for C, C++ and Fortran, respectively. If one needs to drill down through different layers of scripts, then it can be useful to set the compilers using these environment variables. But in general is it recommended to be explicit and use the above CMake cache variables to set the absolute path to the compilers to remove all ambiguity.

If absolute paths to the compilers are not specified using the CMake cache variables or the environment variables as described above, then in MPI mode (i.e. TPL_ENABLE_MPI=ON) TriBITS performs its own search for the MPI compiler wrappers that will find the correct compilers for most MPI distributions (see Configuring with MPI support). However, if in serial mode (i.e. TPL_ENABLE_MPI=OFF), then CMake will do its own default compiler search. The algorithm by which raw CMake finds these compilers is not precisely documented (and seems to change based on the platform). However, on Linux systems, the observed algorithm appears to be:

  1. Search for the C compiler first by looking in PATH (or the equivalent on Windows), starting with a compiler with the name cc and then moving on to other names like gcc, etc. This first compiler found is set to CMAKE_C_COMPILER.
  2. Search for the C++ compiler with names like c++, g++, etc., but restrict the search to the same directory specified by base path to the C compiler given in the variable CMAKE_C_COMPILER. The first compiler that is found is set to CMAKE_CXX_COMPILER.
  3. Search for the Fortran compiler with names like f90, gfortran, etc., but restrict the search to the same directory specified by base path to the C compiler given in the variable CMAKE_C_COMPILER. The first compiler that is found is set to CMAKE_CXX_COMPILER.

WARNING: While this build-in CMake compiler search algorithm may seems reasonable, it fails to find the correct compilers in many cases for a non-MPI serial build. For example, if a newer version of GCC is installed and is put first in PATH, then CMake will fail to find the updated gcc compiler and will instead find the default system cc compiler (usually under /usr/bin/cc on Linux may systems) and will then only look for the C++ and Fortran compilers under that directory. This will fail to find the correct updated compilers because GCC does not install a C compiler named cc! Therefore, if you want to use the default CMake compiler search to find the updated GCC compilers, you can set the CMake cache variable:


or can set the environment variable CC=gcc. Either one of these will result in CMake finding the updated GCC compilers found first in PATH.

Once one has specified the compilers, one can also set the compiler flags, but the way that CMake does this is a little surprising to many people. But the Trilinos TriBITS CMake build system offers the ability to tweak the built-in CMake approach for setting compiler flags. First some background is in order. When CMake creates the object file build command for a given source file, it passes in flags to the compiler in the order:


where <LANG> = C, CXX, or Fortran and <CMAKE_BUILD_TYPE> = DEBUG or RELEASE. Note that the options in CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE> come after and override those in CMAKE_<LANG>_FLAGS! The flags in CMAKE_<LANG>_FLAGS apply to all build types. Optimization, debug, and other build-type-specific flags are set in CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE>. CMake automatically provides a default set of debug and release optimization flags for CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE> (e.g. CMAKE_CXX_FLAGS_DEBUG is typically "-g -O0" while CMAKE_CXX_FLAGS_RELEASE is typically "-O3"). This means that if you try to set the optimization level with -DCMAKE_CXX_FLAGS="-04", then this level gets overridden by the flags specified in CMAKE_<LANG>_FLAGS_BUILD or CMAKE_<LANG>_FLAGS_RELEASE.

Note that TriBITS will set defaults for CMAKE_<LANG>_FLAGS and CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE>, which may be different that what raw CMake would set. TriBITS provides a means for project and package developers and users to set and override these compiler flag variables globally and on a package-by-package basis. Below, the facilities for manipulating compiler flags is described.

Also, to see that the full set of compiler flags one has to actually build a target with, for example make VERBOSE=1 (see Building with verbose output without reconfiguring). One can not just look at the cache variables for CMAKE_<LANG>_FLAGS and CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE> in the file CMakeCache.txt. These get overwritten and redefined by TriBITS in development as described below (see Overriding CMAKE_BUILD_TYPE debug/release compiler options).

The Trilinos TriBITS CMake build system will set up default compile flags for GCC ('GNU') in development mode (i.e. Trilinos_ENABLE_DEVELOPMENT_MODE=ON) on order to help produce portable code. These flags set up strong warning options and enforce language standards. In release mode (i.e. Trilinos_ENABLE_DEVELOPMENT_MODE=ON), these flags are not set. These flags get set internally into the variables CMAKE_<LANG>_FLAGS (when processing packages, not at the global cache variable level) but the user can append flags that override these as described below.

5.4.1   Configuring to build with default debug or release compiler flags

To build a debug version, pass into 'cmake':


This will result in debug flags getting passed to the compiler according to what is set in CMAKE_<LANG>_FLAGS_DEBUG.

To build a release (optimized) version, pass into 'cmake':


This will result in optimized flags getting passed to the compiler according to what is in CMAKE_<LANG>_FLAGS_RELEASE.

The default build type is typically CMAKE_BUILD_TYPE=RELEASE unless -D USE_XSDK_DEFAULTS=TRUE is set in which case the default build type is CMAKE_BUILD_TYPE=DEBUG as per the xSDK configure standard.

5.4.2   Adding arbitrary compiler flags but keeping default build-type flags

To append arbitrary compiler flags to CMAKE_<LANG>_FLAGS (which may be set internally by TriBITS) that apply to all build types, configure with:


where <EXTRA_COMPILER_OPTIONS> are your extra compiler options like "-DSOME_MACRO_TO_DEFINE -funroll-loops". These options will get appended to (i.e. come after) other internally defined compiler option and therefore override them. The options are then pass to the compiler in the order:


This that setting CMAKE_<LANG>_FLAGS can override the default flags that TriBITS will set for CMAKE_<LANG>_FLAGS but will not override flags specified in CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE>.

Instead of directly setting the CMake cache variables CMAKE_<LANG>_FLAGS one can instead set environment variables CFLAGS, CXXFLAGS and FFLAGS for CMAKE_C_FLAGS, CMAKE_CXX_FLAGS and CMAKE_Fortran_FLAGS, respectively.

In addition, if -DUSE_XSDK_DEFAULTS=TRUE is set, then one can also pass in Fortran flags using the environment variable FCFLAGS (raw CMake does not recognize FCFLAGS). But if FFLAGS and FCFLAGS are both set, then they must be the same or a configure error will occur.

Options can also be targeted to a specific TriBITS package using:


The package-specific options get appended to those already in CMAKE_<LANG>_FLAGS and therefore override (but not replace) those set globally in CMAKE_<LANG>_FLAGS (either internally or by the user in the cache).


1) Setting CMAKE_<LANG>_FLAGS will override but will not replace any other internally set flags in CMAKE_<LANG>_FLAGS defined by the Trilinos CMake system because these flags will come after those set internally. To get rid of these project/TriBITS default flags, see below.

2) Given that CMake passes in flags in CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE> after those in CMAKE_<LANG>_FLAGS means that users setting the CMAKE_<LANG>_FLAGS and <TRIBITS_PACKAGE>_<LANG>_FLAGS will not override the flags in CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE> which come after on the compile line. Therefore, setting CMAKE_<LANG>_FLAGS and <TRIBITS_PACKAGE>_<LANG>_FLAGS should only be used for options that will not get overridden by the debug or release compiler flags in CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE>. However, setting CMAKE_<LANG>_FLAGS will work well for adding extra compiler defines (e.g. -DSOMETHING) for example.

WARNING: Any options that you set through the cache variable CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE> will get overridden in the Trilinos CMake system for GNU compilers in development mode so don't try to manually set CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE> directly! To override those options, see CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE>_OVERRIDE below.

5.4.3   Overriding CMAKE_BUILD_TYPE debug/release compiler options

To override the default CMake-set options in CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE>, use:


For example, to default debug options use:


and to override default release options use:

-D CMAKE_C_FLAGS_RELEASE_OVERRIDE="-O3 -funroll-loops" \
-D CMAKE_Fortran_FLAGS_RELEASE_OVERRIDE="-03 -funroll-loops"

NOTES: The TriBITS CMake cache variable CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE>_OVERRIDE is used and not CMAKE_<LANG>_FLAGS_<CMAKE_BUILD_TYPE> because is given a default internally by CMake and the new varaible is needed to make the override explicit.

5.4.5   Turning off strong warnings for individual packages

To turn off strong warnings (for all langauges) for a given TriBITS package, set:


This will only affect the compilation of the sources for <TRIBITS_PACKAGES>, not warnings generated from the header files in downstream packages or client code.

Note that strong warnings are only enabled by default in development mode (Trilinos_ENABLE_DEVELOPMENT_MODE==ON) but not release mode (Trilinos_ENABLE_DEVELOPMENT_MODE==ON). A release of Trilinos should therefore not have strong warning options enabled.

5.4.6   Overriding all (strong warnings and debug/release) compiler options

To override all compiler options, including both strong warning options and debug/release options, configure with:

-D CMAKE_C_FLAGS="-O3 -funroll-loops" \
-D CMAKE_CXX_FLAGS="-03 -fexceptions" \

NOTE: Options like Trilinos_ENABLE_SHADOW_WARNINGS, Trilinos_ENABLE_COVERAGE_TESTING, and Trilinos_ENABLE_CHECKED_STL do not need to be turned off by default but they are shown above to make it clear what other CMake cache variables can add compiler and link arguments.

NOTE: By setting CMAKE_BUILD_TYPE=NONE, then CMAKE_<LANG>_FLAGS_NONE will be empty and therefore the options set in CMAKE_<LANG>_FLAGS will be all that is passed in.

5.4.7   Enable and disable shadowing warnings for all Trilinos packages

To enable shadowing warnings for all Trilinos packages (that don't already have them turned on) then use:


To disable shadowing warnings for all Trilinos packages (even those that have them turned on by default) then use:


NOTE: The default value is empty '' which lets each Trilinos package decide for itself if shadowing warnings will be turned on or off for that package.

5.4.8   Removing warnings as errors for CLEANED packages

To remove the -Werror flag (or some other flag that is set) from being applied to compile CLEANED packages like Teuchos, set the following when configuring:


5.4.9   Adding debug symbols to the build

To get the compiler to add debug symbols to the build, configure with:


This will add -g on most compilers. NOTE: One does not generally need to create a fully debug build to get debug symbols on most compilers.

5.5   Enabling support for C++11

To enable support for C++11 in packages that support C++11 (either optionally or required), configure with:

-D Trilinos_ENABLE_CXX11=ON

By default, the system will try to automatically find compiler flags that will enable C++11 features. If it finds flags that allow a test C++11 program to compile, then it will an additional set of configure-time tests to see if several C++11 features are actually supported by the configured C++ compiler and support will be disabled if all of these features are not supported.

In order to pre-set and/or override the C++11 compiler flags used, set the cache variable:

-D Trilinos_CXX11_FLAGS="<compiler flags>"

5.6   Enabling explicit template instantiation for C++

To enable explicit template instantiation for C++ code for packages that support it, configure with:


When OFF, all packages that have templated C++ code will use implicit template instantiation.

Explicit template instantiation can be enabled (ON) or disabled (OFF) for individual packages with:



For packages that support it, explicit template instantation can massively reduce the compile times for the C++ code involved. To see what packages support explicit instantation just search the CMakeCache.txt file for varibles with ENABLE_EXPLICIT_INSTANTIATION in the name.

5.7   Disabling the Fortran compiler and all Fortran code

To disable the Fortran compiler and all Trilinos code that depends on Fortran set:

-D Trilinos_ENABLE_Fortran=OFF

NOTE: The fortran compiler may be disabled automatically by default on systems like MS Windows.

NOTE: Most Apple Macs do not come with a compatible Fortran compiler by default so you must turn off Fortran if you don't have a compatible Fortran compiler.

5.8   Enabling runtime debug checking

  1. Enabling Trilinos ifdefed runtime debug checking:

To turn on optional ifdefed runtime debug checking, configure with:


This will result in a number of ifdefs to be enabled that will perform a number of runtime checks. Nearly all of the debug checks in Trilinos will get turned on by default by setting this option. This option can be set independent of CMAKE_BUILD_TYPE (which sets the compiler debug/release options).


  • The variable CMAKE_BUILD_TYPE controls what compiler options are passed to the compiler by default while Trilinos_ENABLE_DEBUG controls what defines are set in config.h files that control ifdefed debug checks.
  • Setting -DCMAKE_BUILD_TYPE=DEBUG will automatically set the default Trilinos_ENABLE_DEBUG=ON.
  1. Enabling checked STL implementation:

To turn on the checked STL implementation set:



  • By default, this will set -D_GLIBCXX_DEBUG as a compile option for all C++ code. This only works with GCC currently.
  • This option is disabled by default because to enable it by default can cause runtime segfaults when linked against C++ code that was compiled without -D_GLIBCXX_DEBUG.

5.9   Configuring with MPI support

To enable MPI support you must minimally set:


There is built-in logic to try to find the various MPI components on your system but you can override (or make suggestions) with:

-D MPI_BASE_DIR="path"

(Base path of a standard MPI installation which has the subdirs 'bin', 'libs', 'include' etc.)


-D MPI_BIN_DIR="path1;path2;...;pathn"

which sets the paths where the MPI executables (e.g. mpiCC, mpicc, mpirun, mpiexec) can be found. By default this is set to ${MPI_BASE_DIR}/bin if MPI_BASE_DIR is set.

NOTE: TriBITS uses the MPI compiler wrappers (e.g. mpiCC, mpicc, mpic++, mpif90, etc.) which is more standard with other builds systems for HPC computing using MPI (and the way that MPI implementations were meant to be used). But directly using the MPI compiler wrappers as the direct compilers is inconsistent with the way that the standard CMake module FindMPI.cmake which tries to "unwrap" the compiler wrappers and grab out the raw underlying compilers and the raw compiler and linker command-line arguments. In this way, TriBITS is more consistent with standard usage in the HPC community but is less consistent with CMake (see "HISTORICAL NOTE" below).

There are several different different variations for configuring with MPI support:

  1. Configuring build using MPI compiler wrappers:

The MPI compiler wrappers are turned on by default. There is built-in logic in TriBITS that will try to find the right MPI compiler wrappers. However, you can specifically select them by setting, for example:


which gives the name of the MPI C/C++/Fortran compiler wrapper executable. In this case, just the names of the programs are given and absolute path of the executables will be searched for under ${MPI_BIN_DIR}/ if the cache variable MPI_BIN_DIR is set, or in the default path otherwise. The found programs will then be used to set the cache variables CMAKE_[C,CXX,Fortran]_COMPILER.

One can avoid the search and just use the absolute paths with, for example:

-D MPI_C_COMPILER:FILEPATH=/opt/mpich/bin/mpicc \
-D MPI_CXX_COMPILER:FILEPATH=/opt/mpich/bin/mpic++ \
-D MPI_Fortan_COMPILER:FILEPATH=/opt/mpich/bin/mpif77

However, you can also directly set the variables CMAKE_[C,CXX,Fortran]_COMPILER with, for example:

-D CMAKE_C_COMPILER:FILEPATH=/opt/mpich/bin/mpicc \
-D CMAKE_CXX_COMPILER:FILEPATH=/opt/mpich/bin/mpic++ \
-D CMAKE_Fortan_COMPILER:FILEPATH=/opt/mpich/bin/mpif77

WARNING: If you set just the compiler names and not the absolute paths with CMAKE_<LANG>_COMPILER in MPI mode, then a search will not be done and these will be expected to be in the path at build time. (Note that his is inconsistent the behavior of raw CMake in non-MPI mode described in Selecting compiler and linker options). If both CMAKE_<LANG>_COMPILER and MPI_<LANG>_COMPILER are set, however, then CMAKE_<LANG>_COMPILER will be used and MPI_<LANG>_COMPILER will be ignored.

Note that when USE_XSDK_DEFAULTS=FALSE (see xSDK Configuration Options), then the environment variables CC, CXX and FC are ignored. But when USE_XSDK_DEFAULTS=TRUE and the CMake cache variables CMAKE_[C,CXX,Fortran]_COMPILER are not set, then the environment variables CC, CXX and FC will be used for CMAKE_[C,CXX,Fortran]_COMPILER, even if the CMake cache variables MPI_[C,CXX,Fortran]_COMPILER are set! So if one wants to make sure and set the MPI compilers irrespective of the xSDK mode, then one should set cmake cache variables CMAKE_[C,CXX,Fortran]_COMPILER to the absolute path of the MPI compiler wrappers.

HISTORICAL NOTE: The TriBITS system has its own custom MPI integration support and does not (currently) use the standard CMake module FindMPI.cmake. This custom support for MPI was added to TriBITS in 2008 when it was found the built-in FindMPI.cmake module was not sufficient for the needs of Trilinos and the approach taken by the module (still in use as of CMake 3.4.x) which tries to unwrap the raw compilers and grab the list of include directories, link libraries, etc, was not sufficiently portable for the systems where Trilinos needed to be used. But earlier versions of TriBITS used the FindMPI.cmake module and that is why the CMake cache variables MPI_[C,CXX,Fortran]_COMPILER are defined and still supported.

  1. Configuring to build using raw compilers and flags/libraries:

While using the MPI compiler wrappers as described above is the preferred way to enable support for MPI, you can also just use the raw compilers and then pass in all of the other information that will be used to compile and link your code.

To turn off the MPI compiler wrappers, set:


You will then need to manually pass in the compile and link lines needed to compile and link MPI programs. The compile flags can be set through:


The link and library flags must be set through:


Above, you can pass any type of library or other linker flags in and they will always be the last libraries listed, even after all of the TPLs.

NOTE: A good way to determine the extra compile and link flags for MPI is to use:

export EXTRA_COMPILE_FLAGS="`$MPI_BIN_DIR/mpiCC --showme:compile`"

export EXTRA_LINK_FLAGS="`$MPI_BIN_DIR/mpiCC --showme:link`"

where MPI_BIN_DIR is set to your MPI installations binary directory.

  1. Setting up to run MPI programs:

In order to use the ctest program to run MPI tests, you must set the mpi run command and the options it takes. The built-in logic will try to find the right program and options but you will have to override them in many cases.

MPI test and example executables are passed to CTest ADD_TEST() as:


where <TEST_EXECUTABLE_PATH>, <TEST_ARGS>, and <NP> are specific to the test being run.

The test-independent MPI arguments are:

-D MPI_EXEC:FILEPATH="exec_name"

(The name of the MPI run command (e.g. mpirun, mpiexec) that is used to run the MPI program. This can be just the name of the program in which case the full path will be looked for in ${MPI_BIN_DIR} as described above. If it is an absolute path, it will be used without modification.)


(The default number of processes to use when setting up and running MPI test and example executables. The default is set to '4' and only needs to be changed when needed or desired.)


(The maximum number of processes to allow when setting up and running MPI test and example executables. The default is set to '4' but should be set to the largest number that can be tolerated for the given machine. Tests with more processes than this are excluded from the test suite at configure time.)


(The command-line option just before the number of processes to use <NP>. The default value is based on the name of ${MPI_EXEC}, for example, which is -np for OpenMPI.)

-D MPI_EXEC_PRE_NUMPROCS_FLAGS="arg1;arg2;...;argn"

(Other command-line arguments that must come before the numprocs argument. The default is empty "".)

-D MPI_EXEC_POST_NUMPROCS_FLAGS="arg1;arg2;...;argn"

(Other command-line arguments that must come after the numprocs argument. The default is empty "".)

NOTE: Multiple arguments listed in MPI_EXEC_PRE_NUMPROCS_FLAGS and MPI_EXEC_POST_NUMPROCS_FLAGS must be quoted and seprated by ';' as these variables are interpreted as CMake arrays.

5.10   Configuring for OpenMP support

To enable OpenMP support, one must set:

-D Trilinos_ENABLE_OpenMP=ON

Note that if you enable OpenMP directly through a compiler option (e.g., -fopenmp), you will NOT enable OpenMP inside Trilinos source code.

5.11   Building shared libraries

To configure to build shared libraries, set:


The above option will result in all shared libraries to be build on all systems (i.e., .so on Unix/Linux systems, .dylib on Mac OS X, and .dll on Windows systems).

NOTE: If the project has USE_XSDK_DEFAULTS=ON set, then this will set BUILD_SHARED_LIBS=TRUE by default. Otherwise, the default is BUILD_SHARED_LIBS=FALSE

Many systems support a feature called RPATH when shared libraries are used that embeds the default locations to look for shared libraries when an executable is run. By default on most systems, CMake will automatically add RPATH directories to shared libraries and executables inside of the build directories. This allows running CMake-built executables from inside the build directory without needing to set LD_LIBRARY_PATH on any other environment variables. However, this can be disabled by setting:


but it is hard to find a use case where that would be useful.

5.12   Building static libraries and executables

To build static libraries, turn off the shared library support:


Some machines, such as the Cray XT5, require static executables. To build Trilinos executables as static objects, a number of flags must be set:


The first flag tells cmake to build static versions of the Trilinos libraries. The second flag tells cmake to locate static library versions of any required TPLs. The third flag tells the autodetection routines that search for extra required libraries (such as the mpi library and the gfortran library for gnu compilers) to locate static versions.

NOTE: The flag Trilinos_LINK_SEARCH_START_STATIC is only supported in cmake version 2.8.5 or higher. The variable will be ignored in prior releases of cmake.

5.13   Enabling support for an optional Third-Party Library (TPL)

To enable a given TPL, set:


where <TPLNAME> = BLAS, LAPACK Boost, Netcdf, etc.

The full list of TPLs that is defined and can be enabled is shown by doing a configure with CMake and then grepping the configure output for Final set of .* TPLs. The set of TPL names listed in 'Final set of enabled TPLs' and 'Final set of non-enabled TPLs' gives the full list of TPLs that can be enabled (or disabled).

Some TPLs require only libraries (e.g. Fortran libraries like BLAS or LAPACK), some TPL require only include directories, and some TPLs require both.

Each TPL specification is defined in a FindTPL<TPLNAME>.cmake module file. The job of each of these of these module files is to set the CMake cache variables:

  • TPL_<TPLNAME>_INCLUDE_DIRS:PATH: List of paths to header files for the TPL (if the TPL supplies header files).
  • TPL_<TPLNAME>_LIBRARIES:PATH: List of (absolute) paths to libraries, ordered as they will be on the link line (of the TPL supplies libraries).

These variables are the only variables that are actually used in the CMake build system. Therefore, one can set these two variables as CMake cache variables, for SomeTPL for example, with:

-D TPL_SomeTPL_INCLUDE_DIRS="${LIB_BASE}/include/a;${LIB_BASE}/include/b" \
-D TPL_SomeTPL_LIBRARIES="${LIB_BASE}/lib/;${LIB_BASE}/lib/" \

Using this approach, one can be guaranteed that these libraries and these include directories and will used in the compile and link lines for the packages that depend on this TPL SomeTPL.

WARNING: When specifying TPL_<TPLNAME>_INCLUDE_DIRS and/or TPL_<TPLNAME>_LIBRARIES, the build system will use these without question. It will not check for the existence of these directories or files so make sure that these files and directories exist before these are used in the compiles and links. (This can actually be a feature in rare cases the libraries and header files don't actually get created until after the configure step is complete but before the build step.)

WARNING: Do not try to hack the system and set, for example:

TPL_BLAS_LIBRARIES="-L/some/dir -llib1 -llib2 ..."

This is not compatible with proper CMake usage and it not guaranteed to be supported for all use cases or all platforms! You should instead always use the full library paths when setting TPL_<TPLNAME>_LIBRARIES.

When the variables TPL_<TPLNAME>_INCLUDE_DIRS and TPL_<TPLNAME>_LIBRARIES are not specified, then most FindTPL<TPLNAME>.cmake modules use a default find operation. Some will call FIND_PACKAGE(<TPLNAME>) internally by default and some may implement the default find in some other way. To know for sure, see the documentation for the specific TPL (e.g. looking in the FindTPL<TPLNAME>.cmake file to be sure).

Most TPLs, however, use a standard system for finding include directories and/or libraries based on the function TRIBITS_TPL_FIND_INCLUDE_DIRS_AND_LIBRARIES(). These simple standard FindTPL<TPLNAME>.cmake modules specify a set of header files and/or libraries that must be found. The directories where these header files and library files are looked for are specified using the CMake cache variables:

  • <TPLNAME>_INCLUDE_DIRS:PATH: List of paths to search for header files using FIND_FILE() for each header file, in order.
  • <TPLNAME>_LIBRARY_NAMES:STRING: List of unadorned library names, in the order of the link line. The platform-specific prefixes (e.g.. 'lib') and postfixes (e.g. '.a', '.lib', or '.dll') will be added automatically by CMake. For example, the library, libblas.a, blas.lib or blas.dll will all be found on the proper platform using the name blas.
  • <TPLNAME>_LIBRARY_DIRS:PATH: The list of directories where the library files will be searched for using FIND_LIBRARY(), for each library, in order.

Most FindTPL<TPLNAME>.cmake modules will define a default set of libraries to look for and therefore <TPLNAME>_LIBRARY_NAMES can typically be left off.

In order to allow a TPL that normally requires one or more libraries to ignore the libraries, one can set <TPLNAME>_LIBRARY_NAMES to empty, for example:


Optional package-specific support for a TPL can be turned off by setting:


This gives the user full control over what TPLs are supported by which package independently.

Support for an optional TPL can also be turned on implicitly by setting:


where <TRIBITS_PACKAGE> is a TriBITS package that has an optional dependency on <TPLNAME>. That will result in setting TPL_ENABLE_<TPLNAME>=ON internally (but not set in the cache) if TPL_ENABLE_<TPLNAME>=OFF is not already set.

If all the parts of a TPL are not found on an initial configure the configure will error out with a helpful error message. In that case, one can change the variables <TPLNAME>_INCLUDE_DIRS, <TPLNAME>_LIBRARY_NAMES, and/or <TPLNAME>_LIBRARY_DIRS in order to help fund the parts of the TPL. One can do this over and over until the TPL is found. By reconfiguring, one avoid a complete configure from scrath which saves time. Or, one can avoid the find operations by directly setting TPL_<TPLNAME>_INCLUDE_DIRS and TPL_<TPLNAME>_LIBRARIES.

WARNING: The cmake cache variable TPL_<TPLNAME>_LIBRARY_DIRS does not control where libraries are found. Instead, this variable is set during the find processes and is not actually used in the CMake build system at all.

In summary, this gives the user complete and direct control in specifying exactly what is used in the build process.

TPL Example 1: Standard BLAS Library

Suppose one wants to find the standard BLAS library blas in the directory:


The FindTPLBLAS.cmake module should be set up to automatically find the BLAS TPL by simply enabling BLAS with:


This will result in setting the CMake cache variable TPL_BLAS_LIBRARIES as shown in the CMake output:

-- TPL_BLAS_LIBRARIES='/user/lib/'

(NOTE: The CMake FIND_LIBRARY() command that is used internally will always select the shared library by default if both shared and static libraries are specified, unless told otherwise. See Building static libraries and executables for more details about the handling of shared and static libraries.)

However, suppose one wants to find the blas library in a non-default location, such as in:


In this case, one could simply configure with:

-D BLAS_LIBRARY_DIRS=/projects/something/tpls/lib \

That will result in finding the library shown in the CMake output:

-- TPL_BLAS_LIBRARIES='/projects/something/tpls/'

And if one wants to make sure that this BLAS library is used, then one can just directly set:

-D TPL_BLAS_LIBRARIES=/projects/something/tpls/

TPL Example 2: Intel Math Kernel Library (MKL) for BLAS

There are many cases where the list of libraries specified in the FindTPL<TPLNAME>.cmake module is not correct for the TPL that one wants to use or is present on the system. In this case, one will need to set the CMake cache variable <TPLNAME>_LIBRARY_NAMES to tell the TRIBITS_TPL_FIND_INCLUDE_DIRS_AND_LIBRARIES() function what libraries to search for, and in what order.

For example, the Intel Math Kernel Library (MKL) implementation for the BLAS is usually given in several libraries. The exact set of libraries needed depends on the version of MKL, whether 32bit or 64bit libraries are needed, etc. Figuring out the correct set and ordering of these libraries for a given platform may not be trivial. But once the set and the order of the libraries is known, then one can provide the correct list at configure time.

For example, suppose one wants to use the threaded MKL libraries listed in the directories:


and the list of libraries being searched for is mkl_intel_lp64, mkl_intel_thread, mkl_core and iomp5.

In this case, one could specify this with the following do-configure script:



cmake \
  -D BLAS_LIBRARY_DIRS="${INTEL_DIR}/em64t;${INTEL_DIR}/intel64" \
  -D BLAS_LIBRARY_NAMES="mkl_intel_lp64;mkl_intel_thread;mkl_core;iomp5" \

This would call FIND_LIBRARY() on each of the listed library names in these directories and would find them and list them in:

-- TPL_BLAS_LIBRARIES='/usr/local/intel/Compiler/11.1/064/em64t/;...'

(where ... are the rest of the found libraries.)

NOTE: When shared libraries are used, one typically only needs to list the direct libraries, not the indirect libraries, as the shared libraries are linked to each other.

In this example, one could also play it super safe and manually list out the libraries in the right order by configuring with:


(where ... are the rest of the libraries found in order).

5.14   Disabling support for a Third-Party Library (TPL)

Disabling a TPL explicitly can be done using:


NOTE: If a disabled TPL is a required dependency of some explicitly enabled downstream package, then the configure will error out if Trilinos_DISABLE_ENABLED_FORWARD_DEP_PACKAGES=OFF. Otherwise, a WARNING will be printed and the downstream package will be disabled and configuration will continue.

5.15   Disabling tentatively enabled TPLs

To disable a tentatively enabled TPL, set:


where <TPLNAME> = BinUtils, Boost, etc.

NOTE: Some TPLs in Trilinos are always tentatively enabled (e.g. BinUtils for C++ stacktracing) and if all of the components for the TPL are found (e.g. headers and libraries) then support for the TPL will be enabled, otherwise it will be disabled. This is to allow as much functionality as possible to get automatically enabled without the user having to learn about the TPL, explicitly enable the TPL, and then see if it is supported or not on the given system. However, if the TPL is not supported on a given platform, then it may be better to explicitly disable the TPL (as shown above) so as to avoid the output from the CMake configure process that shows the tentatively enabled TPL being processes and then failing to be enabled. Also, it is possible that the enable process for the TPL may pass, but the TPL may not work correctly on the given platform. In this case, one would also want to explicitly disable the TPL as shown above.

5.16   Disable warnings from TPL header files

To disable warnings coming from included TPL header files for C and C++ code, set:


On some systems and compilers (e.g. GNU), that will result is include directories for all TPLs to be passed in to the compiler using -isystem instead of -I.

WARNING: On some systems this will result in build failures involving gfortran and module files. Therefore, don't enable this if Fortran code in your project is pulling in module files from TPLs.

5.17   xSDK Configuration Options

The configure of Trilinos will adhere to the xSDK configuration standard (todo: put in reference to final document) simply by setting the CMake cache variable:


Setting this will have the following impact:

  • BUILD_SHARED_LIBS will be set to TRUE by default instead of FALSE, which is the default for raw CMake projects (see Building shared libraries).
  • CMAKE_BUILD_TYPE will be set to DEBUG by default instead of RELEASE which is the standard TriBITS default (see CMAKE_BUILD_TYPE).
  • The compilers in MPI mode TPL_ENABLE_MPI=ON or serial mode TPL_ENABLE_MPI=OFF will be read from the environment variables CC, CXX and FC if they are set but the cmake cache variables CMAKE_C_COMPILER, CMAKE_C_COMPILER and CMAKE_C_COMPILER are not set. Otherwise, the TriBITS default behavior is to ignore these environment variables in MPI mode.
  • The Fortran flags will be read from environment variable FCFLAGS if the environment variable FFLAGS and the CMake cache variable CMAKE_Fortran_FLAGS are empty. Otherwise, raw CMake ignores FCFLAGS (see Adding arbitrary compiler flags but keeping default build-type flags).

The rest of the required xSDK configure standard is automatically satisfied by every TriBITS CMake project, including the Trilinos project.

5.18   Generating verbose output

There are several different ways to generate verbose output to debug problems when they occur:

  1. Trace file processing during configure:

This will cause TriBITS to print out a trace for all of the project's, repositorie's, and package's files get processed on lines using the prefix File Trace:. This shows what files get processed and in what order they get processed. To get a clean listing of all the files processed by TriBITS just grep out the lines starting with -- File Trace:. This can be helpful in debugging configure problems without generating too much extra output.

Note that Trilinos_TRACE_FILE_PROCESSING is set to ON automatically when Trilinos_VERBOSE_CONFIGURE = ON.

  1. Getting verbose output from TriBITS configure:

To do a complete debug dump for the TriBITS configure process, use:


However, this produces a lot of output so don't enable this unless you are very desperate. But this level of details can be very useful when debugging configuration problems.

To just view the package and TPL dependencies, it is recommended to use -D Trilinos_DUMP_PACKAGE_DEPENDENCIES = ON.

To just print the link libraries for each library and executable created, use:


Of course Trilinos_DUMP_PACKAGE_DEPENDENCIES and Trilinos_DUMP_LINK_LIBS can be used together. Also, note that Trilinos_DUMP_PACKAGE_DEPENDENCIES and Trilinos_DUMP_LINK_LIBS both default t ON when Trilinos_VERBOSE_CONFIGURE=ON on the first configure.

  1. Getting verbose output from the makefile:

NOTE: It is generally better to just pass in VERBOSE= when directly calling make after configuration is finihsed. See Building with verbose output without reconfiguring.

  1. Getting very verbose output from configure:
-D Trilinos_VERBOSE_CONFIGURE=ON --debug-output --trace

NOTE: This will print a complete stack trace to show exactly where you are.

5.19   Enabling/disabling deprecated warnings

To turn off all deprecated warnings, set:


This will disable, by default, all deprecated warnings in packages in Trilinos. By default, deprecated warnings are enabled.

To enable/disable deprecated warnings for a single Trilinos package, set:


This will override the global behavior set by Trilinos_SHOW_DEPRECATED_WARNINGS for individual package <TRIBITS_PACKAGE>.

5.20   Disabling deprecated code

To actually disable and remove deprecated code from being included in compilation, set:


and a subset of deprecated code will actually be removed from the build. This is to allow testing of downstream client code that might otherwise ignore deprecated warnings. This allows one to certify that a downstream client code is free of calling deprecated code.

To hide deprecated code for a single Trilinos package set:


This will override the global behavior set by Trilinos_HIDE_DEPRECATED_CODE for individual package <TRIBITS_PACKAGE>.

5.21   Outputting package dependency information

To generate the various XML and HTML package dependency files, one can set the output directory when configuring using:


This will generate, by default, the output files TrilinosPackageDependencies.xml, TrilinosPackageDependenciesTable.html, and CDashSubprojectDependencies.xml.

The filepath for TrilinosPackageDependencies.xml can be overridden using:


The filepath for TrilinosPackageDependenciesTable.html can be overridden using:


The filepath for CDashSubprojectDependencies.xml can be overridden using:



  • One must start with a clean CMake cache for all of these defaults to work.
  • The files TrilinosPackageDependenciesTable.html and CDashSubprojectDependencies.xml will only get generated if support for Python is enabled.

5.22   Enabling different test categories

To turn on a set a given set of tests by test category, set:


Valid categories include BASIC, CONTINUOUS, NIGHTLY, HEAVY and PERFORMANCE. BASIC tests get built and run for pre-push testing, CI testing, and nightly testing. CONTINUOUS tests are for post-push testing and nightly testing. NIGHTLY tests are for nightly testing only. HEAVY tests are for more expensive tests that require larger number of MPI processes and longer run times. These test categories are nested (e.g. HEAVY contains all NIGHTLY, NIGHTLY contains all CONTINUOUS and CONTINUOUS contains all BASIC tests). However, PERFORMANCE tests are special category used only for performance testing and don't nest with the other categories.

5.23   Disabling specific tests

Any TriBTS added ctest test (i.e. listed in ctest -N) can be disabled at configure time by setting:

-D <fullTestName>_DISABLE=ON

where <fulltestName> must exactly match the test listed out by ctest -N. Of course specific tests can also be excluded from ctest using the -E argument.

5.24   Trace test addition or exclusion

To see what tests get added and see those that don't get added for various reasons, configure with:


That will print one line per test and will show if the test got added or not. If the test is added, it shows some of the key test properties. If the test did not get added, then this line will show why the test was not added (i.e. due to criteria related to the test's COMM, NUM_MPI_PROCS, CATEGORIES, HOST, XHOST, HOSTTYPE, or XHOSTTYPE arguments).

5.25   Setting test timeouts at configure time

A maximum default time limit (timeout) for all the tests can be set at configure time using the cache variable:


where <maxSeconds> is the number of wall-clock seconds. The default for most projects is 1500 seconds (see the default value set in the CMake cache). This value gets scaled by Trilinos_SCALE_TEST_TIMEOUT and then set as the field TimeOut in the CMake-generated file DartConfiguration.tcl. The value TimeOut from this file is what is directly read by the ctest exectuable. Timeouts for tests are important. For example, when an MPI program has a defect, it can easily hang (forever) until it is manually killed. If killed by a timeout, CTest will kill the test process and all of its child processes correctly.


  • If DART_TESTING_TIMEOUT is not explicitly set by the user, then the projects gives it a default value (typically 1500 seconds but see the value in the CMakeCache.txt file).
  • If DART_TESTING_TIMEOUT is explicitly set to empty (i.e. -DDART_TESTING_TIMEOUT=), then by default tests have no timeout and can run forever until manually killed.
  • Individual tests may have their timeout limit set on a test-by-test basis internally in the project's CMakeLists.txt files (see the TIMEOUT argument for TRIBITS_ADD_TEST() and TRIBITS_ADD_ADVANCED_TEST()). When this is the case, the global timeout set with DART_TESTING_TIMEOUT has no impact on these individually set test timeouts.
  • Be careful not set the global test timeout too low since if a machine becomes loaded tests can take longer to run and may result in timeouts that would not otherwise occur.
  • The value of DART_TESTING_TIMEOUT and the timeouts for individual tests can be scaled up or down using the cache varaible Trilinos_SCALE_TEST_TIMEOUT.
  • To set or override the default global test timeout limit at runtime, see Overriding test timeouts.

5.26   Scaling test timeouts at configure time

The global default test timeout DART_TESTING_TIMEOUT as well as all of the timeouts for the individual tests that have their own timeout set (through the TIMEOUT argument for each individual test) can be scaled by a constant factor <testTimeoutScaleFactor> by configuring with:

-D Trilinos_SCALE_TEST_TIMEOUT=<testTimeoutScaleFactor>

Here, <testTimeoutScaleFactor> can be an integral number like 5 or can be fractional number like 1.5.

This feature is generally used to compensate for slower machines or overloaded test machines and therefore only scaling factors greater than 1 are to be used. The primary use case for this feature is to add large scale factors (e.g. 40 to 100) to compensate for running tests using valgrind (see Running memory checking) but this can also be used for debug-mode builds that create tests which run more slowly than for full release-mode optimized builds.


  • If Trilinos_SCALE_TEST_TIMEOUT is not set, the the default value is set to 1.0 (i.e. no scaling of test timeouts).
  • When scaling the timeouts, the timeout is first truncated to integral seconds so an original timeout like 200.5 will be truncated to 200 before it gets scaled.
  • Only the first fractional digit of Trilinos_SCALE_TEST_TIMEOUT is used so 1.57 is truncated to 1.5, for example, before scaling the test timeouts.
  • The value of the variable DART_TESTING_TIMEOUT is not changed in the CMakeCache.txt file. Only the value of TimeOut written into the DartConfiguration.tcl file (which is directly read by ctest) will be scaled. (This ensures that running configure over and over again will not increase DART_TESTING_TIMEOUT or TimeOut withc each new configure.)

5.27   Enabling support for coverage testing

To turn on support for coverage testing set:


This will set compile and link options -fprofile-arcs -ftest-coverage for GCC. Use 'make dashboard' (see below) to submit coverage results to CDash

5.28   Viewing configure options and documentation

  1. Viewing available configure-time options with documentation:
$ rm -rf CMakeCache.txt CMakeFiles/
$ cmake -LAH -D Trilinos_ENABLE_ALL_PACKAGES=ON \

You can also just look at the text file CMakeCache.txt after configure which gets created in the build directory and has all of the cache variables and documentation.

  1. Viewing available configure-time options without documentation:
$ rm -rf CMakeCache.txt CMakeFiles/
  1. Viewing current values of cache variables:
$ cmake -LA $SOURCE_BASE

or just examine and grep the file CMakeCache.txt.

5.29   Enabling extra repositories with add-on packages:

To configure Trilinos with an post extra set of packages in extra TriBITS repositories, configure with:


Here, <REPOi> is the name of an extra repository that typically has been cloned under the main Trilinos source directory as:


For example, to add the packages from SomeExtraRepo one would configure as:

$ git clone
$ ./do-configure -DTrilinos_EXTRA_REPOSITORIES=SomeExtraRepo \
  [Other Options]

After that, all of the extra packages defined in SomeExtraRepo will appear in the list of official Trilinos packages (after all of the native packages) and one is free to enable any of the defined add-on packages just like any other native Trilinos package.

NOTE: If Trilinos_EXTRAREPOS_FILE and Trilinos_ENABLE_KNOWN_EXTERNAL_REPOS_TYPE are specified, then the list of extra repositories in Trilinos_EXTRA_REPOSITORIES must be a subset and in the same order as the list extra repos read in from the file specified by Trilinos_EXTRAREPOS_FILE. (Also see the variable Trilinos_PRE_REPOSITORIES as well.)

5.30   Enabling extra repositories through a file

In order to provide the list of extra TriBIITS repositories containing add-on packages from a file, configure with:


Specifing extra repositories through an extra repos file allows greater flexibility in the specification of extra repos. This is not helpful for a basic configure of the project but is useful in automated testing using the TribitsCTestDriverCore.cmake script and the script.

The valid values of Trilinos_ENABLE_KNOWN_EXTERNAL_REPOS_TYPE include Continuous, Nightly, and Experimental. Only repositories listed in the file <EXTRAREPOSFILE> that match this type will be included. Note that Nightly matches Continuous and Experimental matches Nightly and Continuous and therefore includes all repos by default.

If Trilinos_IGNORE_MISSING_EXTRA_REPOSITORIES is set to TRUE, then any extra repositories selected who's directory is missing will be ignored. This is useful when the list of extra repos that a given developer develops or tests is variable and one just wants TriBITS to pick up the list of existing repos automatically.

If the file <projectDir>/cmake/ExtraRepositoriesList.cmake exists, then it is used as the default value for Trilinos_EXTRAREPOS_FILE. However, the default value for Trilinos_ENABLE_KNOWN_EXTERNAL_REPOS_TYPE is empty so no extra repostories are defined by default unless Trilinos_ENABLE_KNOWN_EXTERNAL_REPOS_TYPE is specifically set to one of the allowed values.

NOTE: The set of extra repositories listed in the file Trilinos_EXTRAREPOS_FILE can be filtered down by setting the variables Trilinos_PRE_REPOSITORIES if PRE extra repos are listed and/or Trilinos_EXTRA_REPOSITORIES if POST extra repos are listed.

5.31   Reconfiguring completely from scratch

To reconfigure from scratch, one needs to delete the the CMakeCache.txt and base-level CMakeFiles/ directory, for example, as:

$ rm -rf CMakeCache.txt CMakeFiles/
$ ./do-configure [options]

Removing the CMakeCache.txt file is often needed when removing variables from the configure line since they are already in the cache. Removing the CMakeFiles/ directories is needed if there are changes in some CMake modules or the CMake version itself. However, usually removing just the top-level CMakeCache.txt and CMakeFiles/ directory is enough to guarantee a clean reconfigure from a dirty build directory.

If one really wants a clean slate, then try:

$ rm -rf `ls | grep -v do-configure`
$ ./do-configure [options]

WARNING: Later versions of CMake ( require that you remove the top-level CMakeFiles/ directory whenever you remove the CMakeCache.txt file.

5.32   Viewing configure errors

To view various configure errors, read the file:


This file contains detailed output from try-compile commands, Fortran/C name managling determination, and other CMake-specific information.

5.33   Adding configure timers

To add timers to various configure steps, configure with:


This will do baulk timing for the major configure steps which is independent of the number of packages in the project.

To additionally add timing for the configure of individual packages, configure with:


If you are configuring a large number of packages (perhaps by including a lot of add-on packages in extra repos) then you might not want to enable package-by-package timing since it can add some significant overhead to the configure times.

If you just want to time individual packages instead, you can enable that with:



  • This requires that you are running on a Linux/Unix system that has the standard shell command date. CMake does not have built-in timing functions so this system command needs to be used instead. This will report timings to 0.001 seconds but note that the overall configure time will go up due to the increased overhead of calling date as a process shell command.
  • '''WARNING:''' Because this feature has to call the data using CMake's EXECUTE_PROCESS() command, it can be expensive. Therefore, this should really only be turned on for large projects (where the extra overhead is small) or for smaller projects for extra informational purposes.

5.34   Generating export files

The project Trilinos can generate export files for external CMake projects or external Makefile projects. These export files provide the lists of libraries, include directories, compilers and compiler options, etc.

To configure to generate CMake export files for the project, configure with:


This will generate the file TrilinosConfig.cmake for the project and the files <Package>Config.cmake for each enabled package in the build tree. In addition, this will install versions of these files into the install tree.

To confiugre Makefile export files, configure with:


which will generate the file Makefile.export.Trilinos for the project and the files Makefile.export.<Package> for each enabled package in the build tree. In addition, this will install versions of these files into the install tree.

The list of export files generated can be reduced by specifying the exact list of packages the files are requested for with:



  • Only enabled packages will have their export files generated.
  • One would only want to limit the export files generated for very large projects where the cost my be high for doing so.

5.35   Generating a project repo version file

In development mode working with local git repos for the project sources, on can generate a TrilinosRepoVersion.txt file which lists all of the repos and their current versions using:


This will cause a TrilinosRepoVersion.txt file to get created in the binary directory, get installed in the install directory, and get included in the source distribution tarball.

5.36   CMake configure-time development mode and debug checking

To turn off CMake configure-time development-mode checking, set:


This turns off a number of CMake configure-time checks for the Trilinos TriBITS/CMake files including checking the package dependencies. These checks can be expensive and may also not be appropriate for a tarball release of the software. However, this also turns off strong compiler warnings so this is not recommended by default (see <TRIBITS_PACKAGE>_DISABLE_STRONG_WARNINGS). For a release of Trilinos this option is set OFF by default.

One of the CMake configure-time debug-mode checks performed as part of Trilinos_ENABLE_DEVELOPMENT_MODE=ON is to assert the existence of TriBITS package directories. In development mode, the failure to find a package directory is usually a programming error (i.e. a miss-spelled package directory name). But in a tarball release of the project, package directories may be purposefully missing (see Creating a tarball of the source tree) and must be ignored. When building from a reduced tarball created from the development sources, set:


Setting this off will cause the TriBITS CMake configure to simply ignore any missing packages and turn off all dependencies on these missing packages.

Another type of checking is for optional inserted/external packages (e.g. packages who's source can optionally be included in and is flagged with TRIBITS_ALLOW_MISSING_EXTERNAL_PACKAGES()). Any of these package directories that are missing result in the packages being silently ignored by default. However, notes on what missing packages are being ignored can printed by configuring with:


These warnings (starting with 'NOTE', not 'WARNING' that would otherwise trigger warnings in CDash) about missing inserted/external packages will print regardless of the setting for Trilinos_ASSERT_MISSING_PACKAGES.

6   Building (Makefile generator)

This section described building using the default CMake Makefile generator. TriBITS supports other CMake generators such as Visual Studio on Windows, XCode on Macs, and Eclipe project files but using those build systems are not documented here.

6.1   Building all targets

To build all targets use:

$ make [-jN]

where N is the number of processes to use (i.e. 2, 4, 16, etc.) .

6.2   Discovering what targets are available to build

CMake generates Makefiles with a 'help' target! To see the targets at the current directory level type:

$ make help

NOTE: In general, the help target only prints targets in the current directory, not targets in subdirectories. These targets can include object files and all, anything that CMake defines a target for in the current directory. However, running make help it from the base build directory will print all major targets in the project (i.e. libraries, executables, etc.) but not minor targets like object files. Any of the printed targets can be used as a target for make <some-target>. This is super useful for just building a single object file, for example.

6.3   Building all of the targets for a package

To build only the targets for a given TriBITS package, one can use:

$ make <TRIBITS_PACKAGE>_all


$ cd packages/<TRIBITS_PACKAGE>
$ make

This will build only the targets for TriBITS package <TRIBITS_PACKAGE> and its required upstream targets.

6.4   Building all of the libraries for a package

To build only the libraries for given TriBITS package, use:

$ make <TRIBITS_PACKAGE>_libs

6.5   Building all of the libraries for all enabled packages

To build only the libraries for all enabled TriBITS packages, use:

$ make libs

NOTE: This target depends on the <PACKAGE>_libs targets for all of the enabled Trilinos packages. You can also use the target name 'Trilinos_libs.

6.6   Building a single object file

To build just a single object file (i.e. to debug a compile problem), first, look for the target name for the object file build based on the source file, for example for the source file SomeSourceFile.cpp, use:

$ make help | grep SomeSourceFile

The above will return a target name like:

... SomeSourceFile.o

To find the name of the actual object file, do:

$ find . -name "*SomeSourceFile*.o"

that will return something like:


(but this file location and name depends on the source directory structure, the version of CMake, and other factors). Use the returned name (exactly) for the object file returned in the above find operation to remove the object file first, for example, as:

$ rm ./CMakeFiles/<source-dir-path>.dir/SomeSourceFile.cpp.o

and then build it again, for example, with:

$ make SomeSourceFile.o

Again, the names of the target and the object file name an location depend on the CMake version, the structure of your source directories and other factors but the general process of using make help | grep <some-file-base-name> to find the target name and then doing a find find . -name "*<some-file-base-name>*" to find the actual object file path always works.

For this process to work correctly, you must be in the subdirectory where the TRIBITS_ADD_LIBRARY() or TRIBITS_ADD_EXECUTABLE() command is called from its CMakeLists.txt file, otherwise the object file targets will not be listed by make help.

NOTE: CMake does not seem to not check on dependencies when explicitly building object files as shown above so you need to always delete the object file first to make sure that it gets rebuilt correctly.

6.7   Building with verbose output without reconfiguring

One can get CMake to generate verbose make output at build time by just setting the Makefile variable VERBOSE=1, for example, as:


Any number of compile or linking problem can be quickly debugged by seeing the raw compile and link lines. See Building a single object file for more details.

NOTE: The libraries listed on the link line are often in the form -L<lib-dir> -l<lib1> -l<lib2> even if one passed in full library paths for TPLs through TPL_<TPLNAME>_LIBRARIES (see Enabling support for an optional Third-Party Library (TPL)). That is because CMake tries to keep the link lines as short as possible and therefore it often does this translation automatically (whether you want it to or not).

7   Testing with CTest

This section assumes one is using the CMake Makefile generator described above. Also, the ctest does not consider make dependencies when running so the software must be completely built before running ctest as described here.

7.1   Running all tests

To run all of the defined tests (i.e. created using TRIBITS_ADD_TEST() or TRIBITS_ADD_ADVANCED_TEST()) use:

$ ctest -j<N>

(where <N> is an integer for the number of processes to try to run tests in parallel). A summary of what tests are run and their pass/fail status will be printed to the screen. Detailed output about each of the tests is archived in the generate file:


where CTest creates the Testing directory in the local directory where you run it from.

NOTE: The -j<N> argument allows CTest to use more processes to run tests. This will intelligently load ballance the defined tests with multiple processes (i.e. MPI tests) and will try not exceed the number of processes <N>. However, if tests are defined that use more that <N> processes, then CTest will still run the test but will not run any other tests while the limit of <N> processes is exceeded. To exclude tests that require more than <N> processes, set the cache variable MPI_EXEC_MAX_NUMPROCS (see Configuring with MPI support).

7.2   Only running tests for a single package

Tests for just a single TriBITS package can be run with:

$ ctest -j4 -L <TRIBITS_PACKAGE>


$ cd packages/<TRIBITS_PACKAGE>
$ ctest -j4

This will run tests for packages and subpackages inside of the parent package <TRIBITS_PACKAGE>.

NOTE: CTest has a number of ways to filter what tests get run. You can use the test name using -E, you can exclude tests using -I, and there are other approaches as well. See ctest --help and online documentation, and experiment for more details.

7.3   Running a single test with full output to the console

To run just a single test and send detailed output directly to the console, one can run:

$ ctest -R ^<FULL_TEST_NAME>$ -VV

However, when running just a single test, it is usally better to just run the test command manually to allow passing in more options. To see what the actual test command is, use:

$ ctest -R ^<FULL_TEST_NAME>$ -VV -N

This will only print out the test command that ctest runs and show the working directory. To run the test exactly as ctest would, cd into the shown working directory and run the shown command.

7.4   Overriding test timeouts

The configured glboal test timeout described in Setting test timeouts at configure time can be overridden on the CTest command-line as:

$ ctest --timeout <maxSeconds>

This will override the configured cache variable DART_TESTING_TIMEOUT (actually, the scaled value set as TimeOut in the file DartConfiguration.tcl). However, this will not override the test timesouts set on individual tests on a test-by-test basis!

WARNING: Do not try to use --timeout=<maxSeconds> or CTest will just ignore the argument!

7.5   Running memory checking

To configure for running memory testing with valgrind, use:

-D MEMORYCHECK_COMMAND=<abs-path-to-valgrind>/valgrind \
-D MEMORYCHECK_SUPPRESSIONS_FILE=<abs-path-to-supp-file0> \
-D MEMORYCHECK_COMMAND_OPTIONS="-q --trace-children=yes --tool=memcheck \
  --leak-check=yes --workaround-gcc296-bugs=yes \
  --num-callers=50 --suppressions=<abs-path-to-supp-file1> \
  ... --suppressions=<abs-path-to-supp-fileN>"

Above, you have to set the absolute path to the valgrind executable to run using MEMORYCHECK_COMMAND as CMake will not find this for you by default. To use a single valgrind suppression file, just set MEMORYCHECK_SUPPRESSIONS_FILE to the path of that suppression file as shown above. To add other suppression files, they have to be added as other general valgrind arguments in MEMORYCHECK_COMMAND_OPTIONS as shown.

After configuring with the above options, to run the memory tests for all enabled tests, from the base project build directory, do:

$ ctest -T memcheck

This will run valgrind on every test command that is run by ctest.

To run valgrind on the tests for a single package, from the base project directory, do:

$ ctest -T memcheck -L <TRIBITS_PACKAGE>

To run valgrind on a specific test, from the base project directory, do:

$ ctest -T memcheck -R ^<FULL_TEST_NAME>$

Detailed output from valgrind is printed in the file:


NOTE: If you try to run memory tests from any subdirectories, it will not work. You have to run them from the *base project build directory as shown above. A nice way to view valgrind results is to submit to CDash using the dashboard target (see Dashboard submissions).

NOTE: You have to use the valgrind option --trace-children=yes to trace through child processes. This is needed if you have tests that are given as CMake -P scripts (such as advanced tests) or tests driven in bash, Perl, Python, or other languages.

8   Installing

After a build and test of the software is complete, the software can be installed. Actually, to get ready for the install, the install directory must be specified at configure time by setting the variable CMAKE_INSTALL_PREFIX in addition to other variables that affect the installation (see the following sections). The other commands described below can all be run after the build and testing is complete.

For the most typical case where the software is build and installed on the same machine in the same location where it will be used, one just needs to configure with:

$ cmake -DCMAKE_INSTALL_PREFIX=<install-base-dir> [other options] \
$ make -j<N> install

For more details, see the following subsections:

8.1   Setting the install prefix

In order to set up for the install, the install prefix should be set up at configure time by setting, for example:

-D CMAKE_INSTALL_PREFIX=$HOME/install/Trilinos/mpi/opt

The default location for the installation of libraries, headers, and executables is given by the variables (with defaults):

-D Trilinos_INSTALL_INCLUDE_DIR="include" \
-D Trilinos_INSTALL_LIB_DIR="lib" \
-D Trilinos_INSTALL_RUNTIME_DIR="bin" \
-D Trilinos_INSTALL_EXAMPLE_DIR="example"

If these paths are relative (i.e. don't start with "/" and use type STRING) then they are relative to ${CMAKE_INSTALL_PREFIX}. Otherwise the paths can be absolute (use type PATH) and don't have to be under ${CMAKE_INSTALL_PREFIX}. For example, to install each part in any arbitrary location use:

-D Trilinos_INSTALL_INCLUDE_DIR="/usr/Trilinos_include" \
-D Trilinos_INSTALL_LIB_DIR="/usr/Trilinos_lib" \
-D Trilinos_INSTALL_RUNTIME_DIR="/usr/Trilinos_bin" \
-D Trilinos_INSTALL_EXAMPLE_DIR="/usr/share/Trilinos/examples"

NOTE: The defaults for the above include paths will be set by the standard CMake module GNUInstallDirs if Trilinos_USE_GNUINSTALLDIRS=TRUE is set. Some projects have this set by default (see the CMakeCache.txt after configuring to see default being used by this project).

WARNING: To overwrite default relative paths, you must use the data type STRING for the cache variables. If you don't, then CMake will use the current binary directory for the base path. Otherwise, if you want to specify absolute paths, use the data type PATH as shown above.

8.2   Setting install RPATH

Setting RPATH for installed shared libraries and exectuables (i.e. BUILD_SHARED_LIBS=ON) can be a little tricky. Some discussion of how raw CMake handles RPATH and installations can be found at:

The TriBITS/CMake build system being used for this Trilinos CMake project defines the following default behavior for installed RPATH (which is not the same as the raw CMake default behavior):

  • CMAKE_INSTALL_RPATH for all libraries and executables built and installed by this CMake project is set to ${Trilinos_INSTALL_LIB_DIR}. (This default is controlled by the variable Trilinos_SET_INSTALL_RPATH.)
  • The path for all shared external libraries (i.e. TPLs) is set to the location of the external libraries passed in (or automatically discovered) at configure time. (This is controlled by the built-in CMake cache variable CMAKE_INSTALL_RPATH_USE_LINK_PATH which is set to TRUE by default for most TriBITS projects but is empty "" for raw CMake.)

The above default behavior allows the installed exectuables and libraries to be run without needing to set LD_LIBRARY_PATH or any other system environment variables. However, this setting does not allow the installed libraries and executables to be easily moved or relocated. There are several built-in CMake variables that control how RPATH is handled related to installations. The build-in CMake variables that control RPATH handling include CMAKE_INSTALL_RPATH, CMAKE_SKIP_BUILD_RPATH, CMAKE_SKIP_INSTALL_RPATH, CMAKE_SKIP_RPATH, CMAKE_BUILD_WITH_INSTALL_RPATH, CMAKE_INSTALL_RPATH_USE_LINK_PATH. The TriBITS/CMake build system for Trilinos respects all of these raw CMake variables and their documented effect on the build and install.

In addition, this TriBITS/CMake project defines the cache variable:

Trilinos_SET_INSTALL_RPATH: If TRUE, then the global CMake variable CMAKE_INSTALL_RPATH is set to Trilinos_INSTALL_LIB_DIR. If CMAKE_INSTALL_RPATH is set by the user, then that is used instead. This avoids having to manually set CMAKE_INSTALL_RPATH to the correct default install directory.

Rather than re-documenting all of the native CMake RPATH variables mentioned above, instead, we describe how these variables should be set for different installation and distribution scenarios:

  1. Use default CMake behavior
  2. Libraries and executables are built, installed and used on same machine (TriBITS default)
  3. Targets will move after installation
  4. Targets and TPLs will move after installation
  5. Explicitly set RPATH for the final target system
  6. Define all shared library paths at runtime using environment variables

These scenarios in detail are:

  1. Use default CMake behavior: If one just wants the default raw CMake behavior with respect to RPATH, then configure with:


    This will not put any directories into RPATH for the installed libraries or executables. This is the same behavior as setting CMAKE_SKIP_INSTALL_RPATH=TRUE (see Define all shared library paths at runtime using environment variables).

  1. Libraries and executables are built, installed and used on same machine (TriBITS default): One needs no options for this behavior but to make this explicit then configure with:


    As described above, this allows libraries and exectuables to be used right away once installed without needing to set any environment variables.

    Note that this also allows the installed libraries and executables to be moved to the same location on an different but identical machine as well.

  1. Targets will move after installation: In this scenario, the final location of built libraries and executables will be different on the same machine or an otherwise identical machine. In this case, we assume that all of the external library references and directories would be the same. In this case, one would generally configure with:


    Then, to run any exectubles using these shared libraries, one must update LD_LIBRARY_PATH as:

    $ export LD_LIBRARY_PATH=<final-install-dir>/lib:$LD_LIBRARY_PATH

    Or, if the final directory location is known, then one can directly CMAKE_INSTALL_RPATH at configure time to match the final target system and then one does not need to mess with LD_LIBRARY_PATH or any other env variables (see Explicitly set RPATH for the final target system).

  1. Targets and TPLs will move after installation: In this scenario, the final location of the installed libraries and executables will not be the same as the initial install location and the external library (i.e. TPL) locations may not be the same on the final target machine. This can be handled in one of two ways. First, if one knows the final target machine structure, then one can directly set CMAKE_INSTALL_RPATH to the locations on the final target machine (see Explicitly set RPATH for the final target system). Second, if one does not know the final machine directory structure (or the same distribution needs to support several different systems with different directory structures), then one can set CMAKE_SKIP_INSTALL_RPATH=TRUE and then require setting the paths in the env (see Define all shared library paths at runtime using environment variables).
  1. Explicitly set RPATH for the final target system: If one knows the directory structure of the final target machine where the installed libraries and executables will be used, then one can set those paths at configure time with:

    -DCMAKE_INSTALL_RPATH="<path0>;<path1>;..." \

    In this case CMAKE_INSTALL_RPATH is explicitly set. (The value of Trilinos_SET_INSTALL_RPATH has no effect but setting it to FALSE may help to avoid confusion.)

    Once the install directories are moved to the final location, the exectuables can be run without any need to set environment variables.

    Note that TriBITS also accepts the directory separator ":" for:

    -DCMAKE_INSTALL_RPATH="<path0>:<path1>:..." \

    and replaces it internally with ";" which raw CMake requires. (This makes it more robust to pass around inside of CMake code since ";" means array boundary with CMake.). However, since ":" is not a valid character for a path for any Unix system, this is a safe substitution (and CMAKE_INSTALL_RPATH is not used on Windows systems that allow ":" in a directory path).

    Also note that Linux supports RPATHs with the special value $ORIGIN to allow for relative paths and for relocatable installations. (Mac OSX has similar variables like @executable_path.) With this, one can define CMAKE_INSTALL_RPATH using something like $ORIGIN/../lib. See the above CMake RPATH handling reference for more details.

  1. Define all shared library paths at runtime using environment variables: If one wants complete freedom to define the paths for the shared libraries at runtime with LD_LIBRARY_PATH on Linux (and similar variables on other platforms), then one can completely strip RPATH out of the installed libraries and executables by configuring with:


    This will result in all paths being stripped out of RPATH regardless of the values of Trilinos_SET_INSTALL_RPATH or CMAKE_INSTALL_RPATH_USE_LINK_PATH. (This is the same default behavior as raw CMake, see Use default CMake behavior).

    Then the runtime environment must be set up to find the correct shared libraries in the correct order at runtime (e.g. by setting LD_LIBRARY_PATH) . But this approach provides the most flexibility about where executables and libraries are installed and run from.

    Also note that, while not necessary, in order to avoid confusion, it is likely desired to configure with:


    This will produce the least confusing CMake configure output.

One additional issue about RPATH handling on Mac OSX systems needs to be mentioned. That is, in order for this default RPATH approach to work on OSX systems, all of the upstream shared libraries must have @rpath/lib<libname>.dylib embedded into them (as shown by otool -L <lib_or_exec>). For libraries built and installed with CMake, the parent CMake project must be configured with:


For other build systems, see their documentation for shared library support on OSX. To see the proper way to handle RPATH on OSX, just inspect the build and install commands that CMake generates (e.g. using make VERBOSE=1 <target>) for shared libraries and then make sure that these other build systems use equivalent commands. If that is done properly for the chain of all upstream shared libraries then the behaviors of this Trilinos CMake project described above should hold on OSX systems as well.

8.3   Avoiding installing libraries and headers

By default, any libraries and header files defined by in the TriBITS project Trilinos will get installed into the installation directories specified by CMAKE_INSTALL_PREFIX, Trilinos_INSTALL_INCLUDE_DIR and Trilinos_INSTALL_LIB_DIR. However, if the primary desire is to install executables only, then the user can set:


which, if in addition static libraries are being built (i.e. BUILD_SHARED_LIBS=OFF), this this option will result in no libraries or headers being installed into the <install>/include/ and <install>/lib/ directories, respectively. However, if shared libraries are being built (i.e. BUILD_SHARED_LIBS=ON), they the libraries will be installed in <install>/lib/ along with the executables because the executables can't run without the shared libraries being installed.

8.4   Installing the software

To install the software, type:

$ make install

Note that CMake actually puts in the build dependencies for installed targets so in some cases you can just type make -j<N> install and it will also build the software before installing. However, it is advised to always build and test the software first before installing with:

$ make -j<N> && ctest -j<N> && make -j<N> install

This will ensure that everything is built correctly and all tests pass before installing.

WARNING: When using shared libraries, one must be careful to avoid the error "RegularExpression::compile(): Expression too big." when using RPATH and when RPATH gets stripped out on install. To avoid this error, use the shortest build directory you can, like:


This has been shown to allow even the most complex TriBITS-based CMake projects to successfully install and avoid this error.

NOTE: This problem has been resolved in CMake versions 3.6.0+ and does not require a short build directory path.

9   Packaging

Packaged source and binary distributions can also be created using CMake and CPack.

9.1   Creating a tarball of the source tree

To create a source tarball of the project, first configure with the list of desired packages (see Selecting the list of packages to enable) and pass in


To actually generate the distribution files, use:

$ make package_source

The above command will tar up everything in the source tree except for files explicitly excluded in the CMakeLists.txt files and packages that are not enabled so make sure that you start with a totally clean source tree before you do this. You can clean the source tree first to remove all ignored files using:

$ git clean -fd -x

You can include generated files in the tarball, such as Doxygen output files, by creating them first, then running make package_source and they will be included in the distribution (unless there is an internal exclude set).

Disabled subpackages can be included or excluded from the tarball by setting Trilinos_EXCLUDE_DISABLED_SUBPACKAGES_FROM_DISTRIBUTION (the TriBITS project has its own default, check CMakeCache.txt to see what the default is). If Trilinos_EXCLUDE_DISABLED_SUBPACKAGES_FROM_DISTRIBUTION=ON and but one wants to include some subpackages that are otherwise excluded, just enable them or their outer package so they will be included in the source tarball. To get a printout of set regular expresions that will be used to match files to exclude, set:


While a set of default CPack source generator types is defined for this project (see the CMakeCache.txt file), it can be overridden using, for example:


(see CMake documentation to find out the types of supported CPack source generators on your system).

NOTE: When configuring from an untarred source tree that has missing packages, one must configure with:


Otherwise, TriBITS will error out complaining about missing packages. (Note that Trilinos_ASSERT_MISSING_PACKAGES will default to `OFF` in release mode, i.e. Trilinos_ENABLE_DEVELOPMENT_MODE==OFF.)

10   Dashboard submissions

You can use the TriBITS scripting code to submit package-by-package build, test, coverage, memcheck results to the project's CDash dashboard.

First, configure as normal but add the build and test parallel levels with:


(or with some other -j<N>). Then, invoke the build, test and submit with:

$ make dashboard

This invokes the advanced TriBITS CTest scripts to do an experimental build for all of the packages that you have explicitly enabled. The packages that are implicitly enabled due to package dependencies are not directly processed by the experimental_build_test.cmake script.

There are a number of options that you can set in the environment to control what this script does. This set of options can be found by doing:


Currently, the variables can be set include:


For details on what all of these options do, look at the file TribitsCTestDriverCore.cmake (more detailed documentation will come in time). The value of all of these variables is printed out when the make dashboard target is run. From there, one can change and tweak these options. To see the defaults, just run with CTEST_DO_SUBMIT=FALSE and one can see the values used without actually doing the submit. Just grep for the names of the variables in the STDOUT for make dashboard.

For an example of variables one might want to tweak, to run an experimental build and in the process change the build, use:

$ env CTEST_BUILD_NAME=MyBuild make dashboard

After this finishes running, look for the build 'MyBuild' (or whatever build name you used above) in the Trilinos CDash dashboard (the CDash URL is printed at the end of STDOUT).

It is useful to set CTEST_BUILD_NAME to some unique name to make it easier to find your results in the CDash dashboard.

A number of the defaults set in TribitsCTestDriverCore.cmake are overridden from experimental_build_test.cmake (such as CTEST_TEST_TYPE=Experimental) so you will want to look at experimental_build_test.cmake to see how these are changed. The script experimental_build_test.cmake sets reasonable values for these options in order to use the make dashboard target in iterative development for experimental builds.

The target dashboard is not directly related to the built-in CMake targets 'Experimental*' that run standard dashboards with CTest without the custom package-by-package driver in TribitsCTestDriverCore.cmake. The package-by-package extended CTest driver is more appropriate for the TriBITS-based project Trilinos.

Once you configure with -DTrilinos_ENABLE_COVERAGE_TESTING=ON, the environment variable CTEST_DO_COVERAGE_TESTING=TRUE is automatically set by the target 'dashboard' so you don't have to set this yourself.

Doing a memory check with Valgrind requires that you set CTEST_DO_MEMORY_TESTING=TRUE with the 'env' command as:

$ env CTEST_DO_MEMORY_TESTING=TRUE make dashboard

but also note that you may also need to set the valgrind command and options with:

  CTEST_MEMORYCHECK_COMMAND=<abs-path-to-valgrind> \
  CTEST_MEMORYCHECK_COMMAND_OPTIONS="-q --trace-children=yes --tool=memcheck \
   --leak-check=yes --workaround-gcc296-bugs=yes \
   --num-callers=50 --suppressions=<abs-path-to-supp-file1> \
   ... --suppressions=<abs-path-to-supp-fileN>" \
  make dashboard

The CMake cache variable Trilinos_DASHBOARD_CTEST_ARGS can be set on the cmake configure line in order to pass additional arguments to 'ctest -S' when invoking the package-by-package CTest driver. For example:


will set verbose output with CTest.

Also note that one can submit results to a second CDash site as well by setting:

$ env TRIBITS_2ND_CTEST_DROP_SITE=<second-site> \
  TRIBITS_2ND_CTEST_DROP_LOCATION=<second-location> \
  ... \
  make dashboard

If left the same as CTEST_DROP_SITE or CTEST_DROP_LOCATION, then TRIBITS_2ND_CTEST_DROP_SITE and TRIBITS_2ND_CTEST_DROP_LOCATION can be left empty "" and the defaults will be used. However, the user must set at least one of these variables to non-empty in order to trigger the second submit. For example, to submit to an experimental CDash site on the same machine, one would run:

$ env TRIBITS_2ND_CTEST_DROP_LOCATION="/testing/cdash/submit.php?project=Trilinos" \
  ... \
  make dashboard

and TRIBITS_2ND_CTEST_DROP_SITE would be used for CTEST_DROP_SITE. This is a common use case when upgrading to a new CDash installation or testing new features for CDash before impacting the existing CDash site.

Note that if one kills the make dashboard target before it completes, then one must reconfigure from scratch in order to get the build directory back into the same state before the command was run. This is because the dashboard target must first reconfigure the project with no enabled packages before it does the package-by-package configure/build/test/submit which enables each package one at a time. After the package-by-package configure/build/test/submit cycles are complete, then the project is reconfigured with the original set of package enables and returning to the original configure state.