Build and configuration

Configuration

xtensor can be configured via macros which must be defined before including any of its headers. This can be achieved the following ways:

  • either define them in the CMakeLists of your project, with target_compile_definitions cmake command.

  • or create a header where you define all the macros you want and then include the headers you need. Then include this header whenever you need xtensor in your project.

The following macros are already defined in xtensor but can be overwritten:

  • XTENSOR_DEFAULT_DATA_CONTAINER(T, A): defines the type used as the default data container for tensors and arrays. T is the value_type of the container and A its allocator_type.

  • XTENSOR_DEFAULT_SHAPE_CONTAINER(T, EA, SA): defines the type used as the default shape container for tensors and arrays. T is the value_type of the data container, EA its allocator_type, and SA is the allocator_type of the shape container.

  • XTENSOR_DEFAULT_LAYOUT: defines the default layout (row_major, column_major, dynamic) for tensors and arrays. We strongly discourage using this macro, which is provided for testing purpose. Prefer defining alias types on tensor and array containers instead.

  • XTENSOR_DEFAULT_TRAVERSAL: defines the default traversal order (row_major, column_major) for algorithms and iterators on tensors and arrays. We strongly discourage using this macro, which is provided for testing purpose.

The following macros are helpers for debugging, they are not defined by default:

  • XTENSOR_ENABLE_ASSERT: enables assertions in xtensor, such as bound check.

  • XTENSOR_ENABLE_CHECK_DIMENSION: enables the dimensions check in xtensor. Note that this option should not be turned on if you expect operator() to perform broadcasting.

External dependencies

The last group of macros is for using external libraries to achieve maximum performance (see next section for additional requirements):

  • XTENSOR_USE_XSIMD: enables SIMD acceleration in xtensor. This requires that you have xsimd installed on your system.

  • XTENSOR_USE_TBB: enables parallel assignment loop. This requires that you have tbb installed on your system.

  • XTENSOR_DISABLE_EXCEPTIONS: disables c++ exceptions.

  • XTENSOR_USE_OPENMP: enables parallel assignment loop using OpenMP. This requires that OpenMP is available on your system.

Defining these macros in the CMakeLists of your project before searching for xtensor will trigger automatic finding of dependencies, so you don’t have to include the find_package(xsimd) and find_package(TBB) commands in your CMakeLists:

set(XTENSOR_USE_XSIMD 1)
set(XTENSOR_USE_TBB 1)
# xsimd and TBB dependencies are automatically
# searched when the following is executed
find_package(xtensor REQUIRED)

# the target now sets the proper defines (e.g. "XTENSOR_USE_XSIMD")
target_link_libraries(... xtensor)

Build and optimization

Windows

Windows users must activate the /bigobj flag, otherwise it’s almost certain that the compilation fails. More generally, the following options are recommended:

target_link_libraries(... xtensor xtensor::optimize)
set(CMAKE_EXE_LINKER_FLAGS /MANIFEST:NO)

# OR

target_compile_options(target_name PRIVATE /EHsc /MP /bigobj)
set(CMAKE_EXE_LINKER_FLAGS /MANIFEST:NO)

If you defined XTENSOR_USE_XSIMD, you must also specify which instruction set you target:

target_compile_options(target_name PRIVATE /arch:AVX2)
# OR
target_compile_options(target_name PRIVATE /arch:AVX)
# OR
target_compile_options(target_name PRIVATE /arch:ARMv7VE)

If you build on an old system that does not support any of these instruction sets, you don’t have to specify anything, the system will do its best to enable the most recent supported instruction set.

Linux/OSX

Whether you enabled XTENSOR_USE_XSIMD or not, it is highly recommended to build with -march=native option, if your compiler supports it:

target_link_libraries(... xtensor xtensor::optimize)

# OR

target_compile_options(target_name PRIVATE -march=native)

Notice that this option prevents building on a machine and distributing the resulting binary on another machine with a different architecture (i.e. not supporting the same instruction set).