Getting Started¶

Configure and Build¶

MGPU utilizes the CMake build system. A convenient CMake feature are out-of-source builds. MGPU supports this and we encourage you to make use of it.

Browse to the directory where you would like the library to be built and run cmake:

seb@defiant:~$ mkdir mgpubuild
seb@defiant:~$ cd mgpubuild
seb@defiant:~/mgpubuild$ cmake ../mgpu/

CMake will try to configure the library for your system and generate files to build the library as well as unit tests. If for some reason CMake can not locate Boost or CUDA, you may specify the path to those files manually in the CMakeLists.txt file in the MGPU root directory using the following directives:

set(CUDA_TOOLKIT_ROOT_DIR "/path/to/cuda/")
set(BOOST_ROOT "/path/to/boost/")

CMake configures the library for the system it is run on. It determines the number of devices that can be used for computation and their capabilities, in particular if peer-to-peer communicatin is possible between multiple devices. This information is written to the include/mgpu/config.hpp file in the MGPU root folder and can be edited by hand if necessary.

After CMake finished, the library can be built calling calling make mgpu (Linux platform) or nmake mgpu (Windows). CMake can also generate Visual Studio project files if the proper generator is chosen. More information about CMake generators can be found here.

Installing and Using MGPU¶

There are currently no means of installing the MGPU library. To use the library you simply have to link the libmgpu library and add the include path mgpu-root/inlucde.

Testing MGPU¶

The library ships with a self contained collection of unit tests. You may run these to verify that everything works properly on your system. To run the tests browse to the directory where MGPU was built and run ctest:

seb@defiant:~$ cd mgpubuild
seb@defiant:~/mgpubuild$ ctest -j4

The -j4 argument instructs ctest to execute 4 tests in parallel and can speed up the runtime of the tests.