Build Configuration Options¶
RAJA uses BLT, a CMake-based build system. In Getting Started With RAJA, we described how to run CMake to configure RAJA with its default option settings. In this section, we describe RAJA configuration options that are most useful for users to know about and their defaults.
RAJA Option Types¶
Nearly all Cmake options used in RAJA contain the prefix
RAJA_ to give
users flexibility to enable/disable individual compilation features for RAJA,
specifically. RAJA contains two types of options, those that exist in
RAJA only and those that are similar to standard CMake options or options
provided by BLT; i.e., dependent options in CMake terminology. RAJA
dependent option names are the same as the associated CMake and BLT option
names, but with the
RAJA_ prefix added.
RAJA uses a mix of RAJA-only options and CMake-dependent options that can be controlled with CMake or BLT variants.
- Dependent options are typically used for disabling features. For example, providing the option
-DRAJA_ENABLE_TESTS=Offto CMake will disable compilation of RAJA tests, even if the option
-DENABLE_TESTS=Onis also provided.
- We recommend using the option names without the
RAJA_prefix, when available, to enable features at compile time to avoid potential undesired behavior. For example, passing the option
-DRAJA_ENABLE_CUDA=Onto CMake will not enable CUDA because
ENABLE_CUDAis off by default. So to enable CUDA, you need to pass the
-DENABLE_CUDA=Onoption to Cmake.
The RAJA configuration can be set using standard CMake variables along with BLT and RAJA-specific variables. For example, to make a release build with some system default GNU compiler and then install the RAJA header files and libraries in a specific directory location, you could do the following in the top-level RAJA directory:
$ mkdir build-gnu-release $ cd build-gnu-release $ cmake \ -DCMAKE_BUILD_TYPE=Release \ -DCMAKE_C_COMPILER=gcc \ -DCMAKE_CXX_COMPILER=g++ \ -DCMAKE_INSTALL_PREFIX=../install-gnu-release ../ $ make $ make install
Following CMake conventions, RAJA supports three build types:
Debug. With CMake, compiler flags for each of
these build types are applied automatically and so you do not have to
specify them. However, if you want to apply other compiler flags, you will
need to do that using appropriate CMake variables.
All RAJA options are set like regular CMake variables. RAJA settings for
default options, compilers, flags for optimization, etc. can be found in files
RAJA/cmake directory and top-level
Configuration variables can be set by passing
arguments to CMake on the command line when CMake is called, or by setting
options in a CMake cache file and passing that file to CMake using the
-C options. For example, to enable RAJA OpenMP functionality,
pass the following argument to CMake:
The RAJA repository contains a collection of CMake cache files (we call them host-config files) that may be used as a guide for users trying to set their own options. See configopt-raja-hostconfig-label.
Next, we summarize RAJA options and their defaults.
Available RAJA Options and Defaults¶
RAJA uses a variety of custom variables to control how it is compiled. Many
of these are used internally to control RAJA compilation and do
not need to be set by users. Others can be used to enable or disable certain
RAJA features. Most variables get translated to
compiler directives and definitions in the RAJA
config.hpp file that is
generated when CMake runs. The
config.hpp header file is included in other
RAJA headers so all options propagate consistently through the
build process for all of the code.
The following discussion indicates which options exist in RAJA only and those which are dependent options. Dependent options appear in the discussion with parentheses around the ‘RAJA_’ prefix to indicate that the option name without the prefix can be used and is often preferred.
The following tables describe which variables set RAJA options and and their default settings:
- Examples, tests, warnings, etc.
CMake variables can be used to control whether RAJA tests, examples, tutorial exercises, etc. are built when RAJA is compiled.
RAJA can also be configured to build with compiler warnings reported as errors, which may be useful to make sure your application builds cleanly:
Variable Default (RAJA_)ENABLE_WARNINGS_AS_ERRORS Off
RAJA Views/Layouts may be configured to check for out of bounds indexing at runtime:
Variable Default RAJA_ENABLE_BOUNDS_CHECK Off
Note that RAJA bounds checking is a runtime check and will add considerable execution time overhead. Thus, this feature should only be used for correctness checking and should be disabled for production builds.
- Programming model back-end support
Variables that control which RAJA programming model back-ends are enabled are as follows (names are descriptive of what they enable):
Other programming model specific compilation options are also available:
Variable Default (RAJA_)ENABLE_CLANG_CUDA Off (if on, (RAJA_)ENABLE_CUDA must be on too!) RAJA_ENABLE_EXTERNAL_CUB Off RAJA_ENABLE_NV_TOOLS_EXT Off CUDA_ARCH sm_35 (based on hardware support) RAJA_ENABLE_EXTERNAL_ROCPRIM Off RAJA_ENABLE_ROCTX Off RAJA_ENABLE_HIP_INDIRECT_FUNCTION_CALL Off (enables device function pointers in HIP back-end)
(RAJA_)ENABLE_CLANG_CUDA variable on will build CUDA
code with the native support in the Clang compiler.
RAJA_ENABLE_EXTERNAL_CUB variable is used to enable use of an
external install of the NVIDIA CUB support library. When Off, the CUB
library included in the CUDA toolkit will still be used, if available.
Starting with CUDA 11, CUB is installed as part of the CUDA toolkit and
the NVIDIA Thrust library requires that install of CUB. We recommended
projects use the CUB included with the CUDA toolkit for compatibility with
Thrust and applications using Thrust. Users should take note of the CUB
install used by RAJA to ensure they use the same include directories when
configuring their applications.
RAJA_ENABLE_EXTERNAL_ROCPRIM variable is used to enable use of an
external install of the AMD rocPRIM support library. When Off, the
rocPRIM library included in the ROCm install will be used, when available.
We recommend projects use the rocPRIM included with the ROCm install when
available. Users should take note of the rocPRIM install used by RAJA to
ensure they use the same include directories when configuring their
See Getting Started With RAJA for more information about setting other options for RAJA back-ends.
- Data types, sizes, alignment, etc.
RAJA provides type aliases that can be used to parameterize floating point types in applications, which makes it easier to switch between types.
RAJA data types in this section are provided as a convenience to users if they wish to use them. They are not used within RAJA implementation code directly.
The following variables are used to set the data type for the type
Variable Default RAJA_USE_DOUBLE On (type is double) RAJA_USE_FLOAT Off
RAJA::Complex_typecan be enabled to support complex numbers if needed:
Variable Default RAJA_USE_COMPLEX Off
When turned on, the
RAJA::Complex_type is an alias to
There are several variables to control the definition of the RAJA
floating-point data pointer type
RAJA::Real_ptr. The base data type
Real_type. When RAJA is compiled for CPU execution
only, the defaults are:
Variable Default RAJA_USE_BARE_PTR Off RAJA_USE_RESTRICT_PTR On RAJA_USE_RESTRICT_ALIGNED_PTR Off RAJA_USE_PTR_CLASS Off
When RAJA is compiled with CUDA enabled, the defaults are:
Variable Default RAJA_USE_BARE_PTR On RAJA_USE_RESTRICT_PTR Off RAJA_USE_RESTRICT_ALIGNED_PTR Off RAJA_USE_PTR_CLASS Off
The meaning of these variables is:
Variable Meaning RAJA_USE_BARE_PTR Use standard C-style pointer RAJA_USE_RESTRICT_PTR Use C-style pointer with restrict qualifier RAJA_USE_RESTRICT_ALIGNED_PTR Use C-style pointer with restrict qualifier and alignment attribute (see RAJA_DATA_ALIGN below) RAJA_USE_PTR_CLASS Use pointer class with overloaded  operator that applies restrict and alignment intrinsics. This is useful when a compiler does not support attributes in a typedef.
RAJA internally uses a parameter to define platform-specific constant data alignment. The variable that control this is:
Variable Default RAJA_DATA_ALIGN 64
This variable is used to specify data alignment used in intrinsics and typedefs in units of bytes.
For details on the options in this section are used, please see the
- Timer Options
RAJA provides a simple portable timer class that is used in RAJA example codes to determine execution timing and can be used in other apps as well. This timer can use any of three internal timers depending on your preferences, and one should be selected by setting the ‘RAJA_TIMER’ variable.
Variable Values RAJA_TIMER chrono (default) gettime clock
What these variables mean:
Value Meaning chrono Use the std::chrono library from the C++ standard library gettime Use timespec from the C standard library time.h file clock Use clock_t from time.h
- Other RAJA Features
RAJA contains some features that are used mainly for development or may not be of general interest to RAJA users. These are turned off be default. They are described here for reference and completeness.
Variable Meaning RAJA_ENABLE_FT Enable/disable RAJA experimental loop-level fault-tolerance mechanism RAJA_REPORT_FT Enable/disable a report of fault- tolerance enabled run (e.g., number of faults detected, recovered from, recovery overhead, etc.) RAJA_ENABLE_RUNTIME_PLUGINS Enable support for dynamically loaded RAJA plugins. RAJA_ENABLE_DESUL_ATOMICS Replace RAJA atomic implementations with desul variants at compile-time.