This repository contains tutorials for the Kokkos C++ programming model (github.com/kokkos/kokkos).
We are currently running The Kokkos Lectures - an extended version of our Intro-Full Tutorial, spanning 8 modules. For information on that, Slides and Recordings visit: The Kokkos Lectures Wiki
Tutorials in the Intro-Short directory cover
- simple data parallel patterns and policies
- multidimensional array views
- execution and memory spaces
Tutorials in the Intro-Full directory cover
- simple data parallel patterns and policies
- multidimensional array views
- execution and memory spaces
- thread safety and atomic operations
- hierarchical data parallel patterns
All the tutorial folders can be built using either the Makefile
or the CMake CMakeLists.txt
file in each folder.
CMake can build against an installed Kokkos library or download one automatically using FetchContent
.
Without any Kokkos already installed, from an exercise directory, one can run the following:
cmake -B build_dir # -DKokkos_* options
cmake --build build_dir
Kokkos options are described in CMake options.
For example, OpenMP CPU exercises can be built as:
cmake -B build_openmp -DKokkos_ENABLE_OPENMP=ON
cmake --build build_openmp
On Mac, if OpenMP is not available, one can use the Threads backend:
cmake -B build_threads -DKokkos_ENABLE_THREADS=ON
cmake --build build_threads
For a NVIDIA gpu, using gpu arch autodetection:
cmake -B build_cuda -DKokkos_ENABLE_CUDA=ON
cmake --build build_cuda
To pass an already installed Kokkos library, you can use classical CMake variables,
such as Kokkos_ROOT
, or CMAKE_PREFIX_PATH
.
The raw Makefiles require Makefile variables to be properly configured.
In most examples, this is KOKKOS_PATH
pointing to the Kokkos source directory
and KOKKOS_DEVICES
which contains the list of device backends to build.
This will build a new Kokkos library for each exercise.
If you are on a system compatible to our AWS instances, you can type
make
make test
in the Exercises
directory.
Compatible means:
- X86 with a NVIDIA V100 GPU
- kokkos was cloned to ${HOME}/Kokkos/kokkos