title |
---|
Bazel build system |
Drake's primary build system is Bazel. For more information about Bazel, see https://bazel.build/.
Drake also offers a CMake build system wrapper that invokes Bazel under the hood.
Follow Drake's platform-specific setup instructions to install Bazel.
- Start with a git clone of drake, per the Getting Drake instructions.
- Continue with the "Mandatory platform-specific instructions" on the same page.
To build or test Drake, run bazel build or bazel test with the desired target label (and optional configuration options if desired). We give some typical examples below; for more reading about target patterns, see: https://docs.bazel.build/versions/main/user-manual.html#target-patterns.
On Ubuntu, the default compiler is the first gcc
compiler in the PATH
.
On macOS, the default compiler is the Apple LLVM compiler. To use Clang on
Ubuntu, add --config=clang
after any bazel build, bazel test or any
other bazel commands.
Cheat sheet for operating on the entire project:
cd /path/to/drake
bazel build //... # Build the entire project.
bazel test //... # Build and test the entire project.
bazel build --config=clang //... # Build using Clang on Ubuntu.
bazel test --config=clang //... # Build and test using Clang on Ubuntu.
- The "
//
" means "starting from the root of the project". - The "
...
" means "everything including the subdirectories'BUILD
files".- Contrast with, e.g., the "
bazel build common:*
" explained below, where only targets declared directly indrake/common/BUILD
are compiled, and not the targets indrake/common/trajectories/BUILD
. The "*
" matches targets in that directory; the "...
" also matches down into subdirectories.
- Contrast with, e.g., the "
You may use relative pathnames if your shell's working directory is not at the project root:
cd /path/to/drake/common
bazel build ... # Build everything in common and its child subdirectories.
bazel test ... # Test everything in common and its child subdirectories.
bazel build //... # Build the entire project.
bazel test //... # Build and test the entire project.
- As before, the "
...
" above means "everything including subdirectories".- In the first two lines we did not precede "
...
" with "//
", so the search begins in the current directory (common
) and not from thedrake
root. - In the second two lines we used the "
//
" prefix to specify the project root, so we're back to operating on the entire project even thoughcommon
is still our shell's current working directory.
- In the first two lines we did not precede "
Cheat sheet for operating on specific portions of the project:
cd /path/to/drake
bazel build common/... # Build everything in common and its child subdirectories.
bazel build common # Build libcommon.
bazel build common:polynomial # Build libpolynomial.
bazel build common:* # Build everything in common but NOT its children.
bazel test common:polynomial_test # Run one test.
bazel test -c dbg common:polynomial_test # Run one test in debug mode.
bazel test --config=memcheck common:polynomial_test # Run one test under memcheck (valgrind).
bazel test --config=fastmemcheck common:* # Run common's tests under memcheck, with minimal recompiling.
bazel test --config=kcov common:polynomial_test # Run one test under kcov (see instructions below).
bazel build -c dbg common:polynomial_test && \
gdb bazel-bin/common/polynomial_test # Run one test under gdb.
bazel test -c dbg --config=clang --config=asan common:polynomial_test # Run one test under AddressSanitizer.
bazel test --config lint //... # Only run style checks; don't build or test anything else.
- The "
:
" syntax separates target names from the directory path of theBUILD
file they appear in. In this case, for example,drake/common/BUILD
specifiescc_test(name = "polynomial_test")
. - Note that the configuration switches (
-c
and--config
) influence the entire command. For example, running a test indbg
mode means that its prerequisite libraries are also compiled and linked indbg
mode. - For the definitions of the "
--config
" options seedrake/tools/bazel.rc
.
In general, to figure out what binary-specific arguments are available, add
"-- --help
" to your bazel run
command. An an example,
bazel run //examples/acrobot:run_passive -- --help
The bare --
separates Bazel arguments from the program's arguments.
For running tests, you may pass custom arguments to the test program via --test_arg.
For a C++ unittest that uses drake_cc_googletest
, for example:
bazel test multibody/plant:multibody_plant_test --test_output=streamed --nocache_test_results --test_arg=--gtest_filter='*SimpleModelCreation*'
For a Python unittest that uses drake_py_unittest
, for example:
bazel test bindings/pydrake:py/symbolic_test --test_output=streamed --nocache_test_results --test_arg=--trace=user --test_arg=TestSymbolicVariable
On macOS, DWARF debug symbols are emitted to a .dSYM
file. The Bazel
cc_binary
and cc_test
rules do not natively generate or expose this
file, so we have implemented a workaround in Drake, --config=apple_debug
.
This config turns off sandboxing, which allows a genrule
to access the
.o
files and process them into a .dSYM
. Use as follows:
bazel build --config=apple_debug path/to/my:binary_or_test_dsym
lldb ./bazel-bin/path/to/my/binary_or_test
For more information, see bazelbuild/bazel#2537.
Please use the "buildifier
" tool to format edits to BUILD
files (in the
same spirit as clang-format
formatting C++ code):
cd /path/to/drake
bazel-bin/tools/lint/buildifier --all # Reformat all Bazel files.
bazel-bin/tools/lint/buildifier common/BUILD # Only reformat one file.
In most cases the bazel-bin/tools/lint/buildifier
will already be compiled
by the time you need it. In case it's absent, you can compile it via:
cd /path/to/drake
bazel build //tools/lint:buildifier
The Drake Bazel build currently supports the following proprietary solvers:
- Gurobi 9.5
- MOSEK™ 10.0
- SNOPT 7.4
- Register for an account on https://www.gurobi.com.
- Set up your Gurobi license file in accordance with Gurobi documentation.
export GRB_LICENSE_FILE=/path/to/gurobi.lic
.- Download
gurobi9.5.1_linux64.tar.gz
. You may need to manually edit the URL to get the correct version. - Unzip it. We suggest that you use
/opt/gurobi951
to simplify working with Drake installations. - If you unzipped into a location other than
/opt/gurobi951
, then callexport GUROBI_HOME=GUROBI_UNZIP_PATH/linux64
to set the path you used, where inGUROBI_HOME
folder you can findbin
folder.
- Register for an account on http://www.gurobi.com.
- Set up your Gurobi license file in accordance with Gurobi documentation.
export GRB_LICENSE_FILE=/path/to/gurobi.lic
- Download and install
gurobi9.5.1_mac64.pkg
.
To confirm that your setup was successful, run the tests that require Gurobi:
bazel test --config gurobi --test_tag_filters=gurobi //...
The default value of --test_tag_filters
in Drake's bazel.rc
excludes
these tests. If you will be developing with Gurobi regularly, you may wish
to specify a more convenient --test_tag_filters
in a local .bazelrc
.
See https://docs.bazel.build/versions/main/user-manual.html#bazelrc.
The Drake Bazel build system downloads MOSEK™ 10.0.18 automatically. No manual installation is required. Set the location of your license file as follows:
export MOSEKLM_LICENSE_FILE=/path/to/mosek.lic
To confirm that your setup was successful, run the tests that require MOSEK™:
bazel test --config mosek --test_tag_filters=mosek //...
The default value of --test_tag_filters
in Drake's bazel.rc
excludes
these tests. If you will be developing with MOSEK™ regularly, you may wish
to specify a more convenient --test_tag_filters
in a local .bazelrc
.
See https://docs.bazel.build/versions/main/user-manual.html#bazelrc.
Drake provides two mechanisms to include the SNOPT sources. One mechanism is to provide your own SNOPT source archive. The other mechanism is via access to a private RobotLocomotion git repository.
- Download the SNOPT sources from the distributor in
.tar.gz
format (e.g., namedsnopt7.4.tar.gz
). export SNOPT_PATH=/home/username/Downloads/snopt7.4.tar.gz
Using the RobotLocomotion git repository
- Obtain access to the private RobotLocomotion/snopt GitHub repository.
- Set up SSH access to github.com.
export SNOPT_PATH=git
Test the build (for either mechanism)
To confirm that your setup was successful, run the tests that require SNOPT:
bazel test --config snopt --test_tag_filters=snopt //...
The default value of --test_tag_filters
in Drake's bazel.rc
excludes
these tests. If you will be developing with SNOPT regularly, you may wish
to specify a more convenient --test_tag_filters
in a local .bazelrc
.
See https://docs.bazel.build/versions/main/user-manual.html#bazelrc.
SNOPT support has some known problems on certain programs (see drake issue #10422 for a summary).
Drake is in the process of adding support for multiprocessing using OpenMP. At the moment, that support is experimental and is not recommended for Drake's users.
For Drake Developers who wish to enable OpenMP, use this config switch:
bazel test --config omp //...
This switch is enabled in CI under the "Ubuntu Everything" build flavor.
The Drake Bazel build system has integration support for some optional development tools:
- kcov -- test coverage analysis
kcov
can analyze coverage for any binary that contains DWARF format
debugging symbols, and produce nicely formatted browse-able coverage reports.
Drake's kcov
build system integration is only supported on Ubuntu, not
macOS.
In some cases, running kcov builds and regular builds from the same source
tree will lead to Bazel error messages like "this rule is missing dependency
declarations". To resolve that problem, either run the kcov build from a
fresh checkout, or else run a bazel clean
.
To analyze test coverage, run one (or more) tests under kcov
:
bazel test --config=kcov common:polynomial_test
Note that it disables compiler-optimization (-O0
) to have a better and more
precise coverage report. If you have trouble with kcov and unoptimized programs,
you can turn it back on by also supplying --copt -O2
.
For each test program, individual coverage reports are written to per-target
directories. Use the kcov_tool
to merge coverage data into a new directory:
tools/dynamic_analysis/kcov_tool merge [OUTPUT-DIR]
To view the merged data, browse to index.html
in the OUTPUT-DIR.
In a local developer workspace, coverage data may accumulate over successive
build jobs, even if source files or other dependencies have changed. The stale
data would be scattered within the directory tree linked as
bazel-testlogs
. To clear out old data, use kcov_tool clean
:
tools/dynamic_analysis/kcov_tool clean
Coverage reports for Python sources produced by kcov are useful, but can be misleading. As of Ubuntu 20.04 and kcov 38, Python reports do not render coverage for multi-line statements properly. Statements that use delimiter pairs to span more than two lines, or statements that use string token pasting across multiple lines may be mistakenly shown as only partially executed.