The purpose of these test programs is to experiment with various heterogeneous frameworks and "performance portability" approaches.
| Test | CPU (serial) | CUDA | Alpaka/Cupla | SYCL |
|---|---|---|---|---|
vector_add |
✔️ | ✔️ | ✔️ ✔️ ✔️ ✔️ | ✔️ ✔️ ✔️ |
shared_mem |
✔️ | ✔️ | ✔️ ✔️ ✔️ ✔️ | ✔️ ✔️ ✔️ |
Support for Alpaka/Cupla is split in
- CPU serial
- CPU paraller, using OpenMP
- CPU paraller, using TBB
- GPU parallel, using CUDA
Support for SYCL is split in
- CPU (unspecified)
- generic OpenCL
- NVIDIA PTX
These programs require a recent CUDA version (nvcc supporting C++14 and --expt-relaxed-constexpr) and a machine with GPU.
Alpaka/Cupla support different backends; so far it has been tested with serial, TBB and OpenMP backends for the CPU, and the CUDA backend for NVIDIA GPUs. The latter requires CUDA 9.2 through 10.1, and has been tested with gcc 7.x and gcc 8.x.
Rather than using the advertised CMake-based approach, one can use Cupla as header-only library:
# choose a base directory for Alpaka and Cupla
BASE=/opt/alpaka
mkdir -p $BASE
cd $BASE
# download the development version of Alpaka
git clone -b develop git@github.com:ComputationalRadiationPhysics/alpaka.git
# download the development version of Capla, and set up sparse checkout to remove the embedded Alpaka
git clone -b dev git@github.com:ComputationalRadiationPhysics/cupla.git
cd cupla
git config core.sparseCheckout true
echo -e '/*\n!/alpaka/' > .git/info/sparse-checkout
git read-tree -mu HEAD
Note: these instructions were accurate as of October 15th, 2019.
See the vector_add/Makefile for an example.
If you use a different BASE folder, updated the Makefiles accordingly.
These programs were tested with CodePlay's ComputeCpp Community Edition, version 1.1.4 .
Other projects with similar goals:
- BabelStream memory transfer benchmark: https://uob-hpc.github.io/BabelStream/
- Parallel Research Tools: https://github.com/ParRes/Kernels
- DoE portability study: https://performanceportability.org/