Installation

Installation#

Introduction#

This chapter describes how to obtain rocRAND. There are two main methods: the easiest way is to install the prebuilt packages from the ROCm repositories. Alternatively, this chapter also describes how to build rocRAND from source.

Prebuilt packages#

Installing the prebuilt rocRAND packages requires a ROCm-enabled platform. See the ROCm documentation for more information. After installing ROCm or enabling the ROCm repositories, rocRAND can be obtained using the system package manager.

For Ubuntu and Debian:

sudo apt-get install rocrand

For CentOS:

sudo yum install rocrand

For SLES:

sudo dnf install rocrand

This will install rocRAND into the /opt/rocm directory.

Building rocRAND from source#

Obtaining sources#

The rocRAND sources are available from the rocRAND GitHub Repository. Use the branch that matches the system-installed version of ROCm. For example on a system that has ROCm 5.3 installed, use the following command to obtain rocRAND sources:

git checkout -b rocm-5.3 https://github.com/ROCmSoftwarePlatform/rocRAND.git

Building the library#

After obtaining the sources, rocRAND can be built using the installation script:

cd rocRAND
./install --install

This automatically builds all required dependencies, excluding HIP and Git, and installs the project to /opt/rocm if everything went well. See ./install --help for further information.

Building with CMake#

For a more elaborate installation process, rocRAND can be built manually using CMake. This enables certain configuration options that are not exposed to the ./install script. In general, rocRAND can be built using CMake by configuring as follows:

cd rocrand; mkdir build; cd build
# Configure the project
CXX=<compiler> cmake [options] ..
# Build
make -j4
# Optionally, run the tests
ctest --output-on-failure
# Install
[sudo] make install

To build for the ROCm platform,``<compiler>`` should be set to hipcc. When building for CUDA or HIP-CPU, <compiler> should be set to the host compiler. If building for CUDA, then the location of nvcc may need to be passed explicitly using -DCMAKE_CUDA_COMPILER=<path-to-nvcc> if it is not on the path. Additionally, the directory where FindHIP.cmake is installed needs to be passed explicitly using -DCMAKE_MODULE_PATH. By default, this file is installed in /opt/rocm/hip/cmake.

The following configuration options are available, in addition to the built-in CMake options:

  • BUILD_FORTRAN_WRAPPER controls whether to build the Fortran wrapper. Defaults to OFF.

  • BUILD_TEST controls whether to build the rocRAND tests. Defaults to OFF.

  • BUILD_BENCHMARK controls whether to build the rocRAND benchmarks. Defaults to OFF.

  • BUILD_ADDRESS_SANITIZER controls whether to build with address sanitization enabled. Defaults to OFF.

  • USE_HIP_CPU is an experimental option that controls whether to build for HIP-CPU. Defaults to OFF.

To install rocRAND with a non-standard installation location of ROCm, pass -DCMAKE_PREFIX_PATH=</path/to/opt/rocm/> or set the environment variable ROCM_PATH to path/to/opt/rocm.

Building the Python API wrapper#

Requirements#

The rocRAND Python API Wrapper requires the following dependencies:

  • rocRAND

  • Python 3.5

  • NumPy (will be installed automatically as a dependency if necessary)

Note: If rocRAND is built from sources but not installed or installed in non-standard directory, set the ROCRAND_PATH environment variable. For example:

export ROCRAND_PATH=~/rocRAND/build/library/

Installing#

The Python rocRAND module can be installed using pip:

cd rocrand/python/rocrand
pip install .

The tests can be executed as follows:

cd rocrand/python/rocrand
python tests/rocrand_test.py
Contents