Skip to content

Install SELF#

Install with Spack#

The easiest way to get started is to use the spack package manager. On a Linux platform, set up spack

git clone https://github.com/spack/spack ~/spack
source ~/spack/share/spack/setup-env.sh

Allow spack to locate your compilers (make sure you have C, C++, and Fortran compilers installed!)

spack compiler find

SELF comes with a spack environment file that defines the dependencies that are required for SELF. The versions listed in this environment file are the specific versions we regularly test against. To get this environment file, clone the SELF repository

git clone https://github.com/fluidnumerics/SELF/ ~/SELF/

If you have a preferred compiler you would like for spack to use, you can use spack config add, e.g.

spack -e ~/SELF/share/spack-env config add packages:all:require:["'%gcc@12.2.0'"]

The example above will force packages to be built with version 12.2.0 of gfortran from the gcc compiler set.

Note

If you do not set a preferred compiler, spack will pick one based on the available compilers found using spack compiler find

To reduce build time, import existing packages on your system 

spack external find --not-buildable

Next, install SELF's dependencies (OpenMPI, HDF5, and feq-parse) 

spack -e ~/SELF/share/spack-env install --no-check-signature

Then, install SELF 

cd ~/SELF
spack env activate ~/SELF/share/spack-env
mkdir ~/SELF/build
cd ~/SELF/build
cmake -DCMAKE_INSTALL_PREFIX=${HOME}/opt/self ../
make
make install

If you'd like to run the tests included with SELF, to verify your installation, you can use ctest.

cd ${HOME}/opt/self/test
ctest

Once v0.0.1 is released#

The easiest way to get started is to use the spack package manager. The spack package manager provides you with an easy command line interface to install research software from source code with all of its dependencies.

Note

Before proceeding, you will need to ensure that you have a 2008 compliant Fortran compiler.

On a Linux platform, set up spack : 

git clone https://github.com/spack/spack ~/spack
source ~/spack/share/spack/setup-env.sh

Allow spack to locate your compilers (make sure you have C, C++, and Fortran compilers installed!)

spack compiler find

The example above will force packages to be built with version 12.2.0 of gfortran from the gcc compiler set.

To reduce build time, import existing packages on your system

spack external find --not-buildable

Next, install SELF and it's dependencies

spack install self

By default, this will install SELF with the following features * Double precision floating point arithmetic * No unit tests and no examples * No multi-threading, CPU-only

You can view documentation on all possible variants using

spack info self

Enable Multithreading#

Many of the computationally intensive methods in SELF are written using the do concurrent structure. We have provided the variant +multithreading which will enable multithreading for all do concurrent blocks. You can install SELF with multithreading using

spack install self+multithreading

If you are using the GNU compiler suite, the number of threads used for do concurrent blocks is determined during build time. Because of this, we have provided the nthreads option, which defaults to 4. You can change this option to a value more sensible for your platform, e.g.

spack install self+multithreading nthreads=16 % gcc

The %gcc here indicates that you intend to build SELF with the GNU compilers.

Enable Nvidia GPU Acceleration#

SELF provides GPU accelerated implementations of all methods that are used in forward stepping conservation law solvers. On Nvidia GPU platforms, you can take advantage of this using the +cuda variant : 

spack install self+cuda
This will also ensure that the MPI flavor that is used is GPU aware. You can specify the GPU architecture using the gpu_arch build option, e.g. for A100 GPUs

spack install self+cuda gpu_arch=sm_80

Note

AMD GPU-Aware MPI is currently not available in Spack. This means that these steps will not allow you to build SELF for multi-GPU platforms with AMD GPUs. See Advanced Installation for details on how to install for AMD GPU platforms.

Advanced Installation#

Dependencies#

The Spectral Element Library in Fortran can be built provided the following dependencies are met

Learn more about installing SELF's dependencies

Installing SELF from source with CMake#

Warning

It is assumed that you have all of the necessary dependencies installed on your system and that they are discoverable by CMake.

SELF comes with a CMake build system that defines the build and installation process.

  1. Clone the SELF repository
git clone https://github.com/fluidnumerics/self ~/self/
  1. Create a build directory
mkdir ~/self/build
cd ~/self/build
  1. Run CMake to build the make system. You can set CMAKE_INSTALL_PREFIX to a path where you'd prefer to install SELF; here, we set it to ${HOME}/opt/self
cmake -DCMAKE_INSTALL_PREFIX=${HOME}/opt/self ../
  1. Build SELF and run the test suite to ensure everything is built properly
make
ctest --output-on-failure
  1. Install SELF 
make install

When you install SELF, you will install the following artifacts

  • ${CMAKE_INSTALL_PREFIX}/lib/libself-static.a - A static library for SELF
  • ${CMAKE_INSTALL_PREFIX}/lib/libself.so - A shared object library for SELF
  • ${CMAKE_INSTALL_PREFIX}/include/*.mod - Module files generated by the Fortran compiler during the build process.
  • ${CMAKE_INSTALL_PREFIX}/example/* - A set of example programs
  • ${CMAKE_INSTALL_PREFIX}/test/* - A set of unit tests

By default, this will install SELF with the following features * Double precision floating point arithmetic * No unit tests and no examples * No multi-threading, CPU-only

There are a few CMake options that you can set to control the build features : * SELF_ENABLE_MULTITHREADING: Option to enable CPU multithreading for do concurrent loop blocks. (Default: OFF) * SELF_ENABLE_TESTING: Option to enable build of tests. (Default: ON) * SELF_ENABLE_EXAMPLES: Option to enable build of examples. (Default: ON) * SELF_ENABLE_GPU: Option to enable GPU backend. Requires either CUDA or HIP. (Default: OFF) * SELF_ENABLE_DOUBLE_PRECISION Option to enable double precision for floating point arithmetic. (Default: ON)

Enabling Multithreading CPU support#

Computationally heavy methods in SELF are expressed using Fortran's do concurrent loop blocks, which gives compilers the freedom to parallelize operations. Every Fortran compiler has their own set of compiler flags to enable parallelization of do concurrent blocks (see this post on the Fortran-Lang discourse). We have provided a single option in the CMake build system that allow you to enable parallelization. At the cmake stage of the build process, you can set SELF_ENABLE_MULTITHREADING=ON, e.g.

cmake -DSELF_ENABLE_MULTITHREADING=ON \
      -DCMAKE_INSTALL_PREFIX=${HOME}/opt/self \
       ../

If you are building with the GNU compilers (gfortran), the number of threads used for parallelization is determined at build time. By default, the SELF build system will set the number of threads to 4. You can override this setting at the cmake stage of the build process using the SELF_MULTITHREADING_NTHREADS build variable, e.g. to set the number of threads to 8 with gfortran,

cmake -DSELF_ENABLE_MULTITHREADING=ON \
      -DSELF_MULTITHREADING_NTHREADS=8 \
      -DCMAKE_INSTALL_PREFIX=${HOME}/opt/self \
       ../

The CMake build system will set the appropriate flags for multithreading for GNU, Intel (ifort and ifx), LLVM, and Nvidia HPC Compilers. If you are not using gfortran, you can set the number of threads for parallelism at runtime using the OMP_NUM_THREADS environment variable

Enabling GPU Support#

SELF offers the option to use HIP or CUDA. Some of our "heavy-lifting" kernels, such as divergence, gradient, and grid interpolation operations are expressed using the BLAS API. For these, we use HIPBLAS or CUBLAS. GPU support is enabled in the CMake stage of the build by setting SELF_ENABLE_GPU=ON

The CMake build system will automatically search for HIP. If HIP is not found, then it will search for CUDA. If neither is found, the build process will fail.

HIP#

HIP can be used to build for either AMD or Nvidia GPU's. If you have HIP installed and it is found, you can also set the CMAKE_HIP_ARCHITECTURES build variable to specify which GPU architecture you want to build for. Alternatively, if you are building SELF on the system that has a GPU installed, you can let HIP auto-detect the available GPU. At this time, we also advise setting the CXX enviornment variable to hipcc,e.g.

CXX=hipcc \
cmake -DSELF_ENABLE_GPU=ON \
      -DCMAKE_INSTALL_PREFIX=${HOME}/opt/self \
       ../

CUDA#

SELF provides you the ability to use CUDA directly, in case you are on a system that does not have AMD's ROCm and HIP installed. If you have CUDA installed and it is found, you can also set the CMAKE_CUDA_ARCHITECTURES build variable to specify which Nvidia GPU architecture you want to build for. At this time, we also advise setting the CXX enviornment variable to nvcc,e.g.

CXX=nvcc \
cmake -DSELF_ENABLE_GPU=ON \
      -DCMAKE_INSTALL_PREFIX=${HOME}/opt/self \
       ../

Reference table for GPU architecture codes#

Vendor Model Architecture code(s)
AMD Instinct MI100 gfx908
AMD Instinct MI210 gfx90a
AMD Instinct MI250 gfx90a
AMD Instinct MI250x gfx90a
AMD Radeon Pro W7900 gfx1100
AMD Radeon Pro W7800 gfx1100
Nvidia Volta (V100) sm_70, sm_72
Nvidia Ampere (A100) sm_80, sm_86, sm_87
Nvidia Hopper (H100) sm_90, sm_90a

If you encounter any problems, feel free to open an new issue