SYCL Container

This container contains an environment to run and compile SYCL 2020 codes on NVIDIA and AMD GPUs and any CPU with two SYCL compilers:

• AdaptiveCpp v24.02.0
• oneAPI DPC++ (Intel LLVM release 2023-WW27)

How to run container

Singularity

• Documentation for GPU Support

singularity pull docker://ghcr.io/maison-de-la-simulation/sycl-complete

#optional: Mount directory
export SINGULARITY_BIND="/path/to/local/dir:/path/to/container/dir"

#Run container in shell
singularity shell --env OMP_NUM_THREADS=64 --env OMP_PLACES=cores sycl-complete.sif
singularity shell --nv sycl-complete.sif

#Or execute command inside container
singularity exec --nv   sycl-complete.sif acpp-info
singularity exec --rocm sycl-complete.sif sycl-ls

Docker

docker pull ghcr.io/maison-de-la-simulation/sycl-complete

#Run container in interactive mode
docker run --it ghcr.io/maison-de-la-simulation/sycl-complete
..

Use the SYCL compilers & runtime

Compilers are installed in /opt/sycl. The environment is setup in the $PATH so just type in sycl-ls (DPC++) or acpp-info (AdaptiveCpp) to list the SYCL devices. The associated runtimes are installed so once your code is compiled you should be able to run it within the container.

Following are examples to compile a SYCL code with the two different SYCL compilers. We assume the code is mounted inside /mnt/program.

AdaptiveCpp

• Documentation for using acpp.

With AdaptiveCpp, set the ACPP_TARGETS before using the syclcc compiler.

#We build for a NVIDIA A100, AMD MI250x, and GPUs
export ACPP_TARGETS="omp;cuda:sm_80;hip:gfx90a"

cd /mnt/program/build #go into build folder

#use /opt/sycl/acpp/syclcc as CXX compiler
cmake -DCMAKE_CXX_COMPILER=syclcc ..
make -j 64

DPC++

• Intel DPC++ Users manual

With this compiler, use the -fsycl and -fsycl-targets compilation flags to use ahead of time compilation. We use the cmake M_FLAGS variable to pass these parameters to the clang++ compiler.

#for NVIDIA A100
cmake \
-DCMAKE_CXX_COMPILER=clang++ \
-DM_FLAGS="-fsycl -fsycl-targets=nvidia_gpu_sm_80" \
..


#for Intel CPU with avx512 capabilities
clang++ -fsycl -fsycl-targets=spir64_x86_64 -Xsycl-target-backend "-march=avx512" myprog.cpp myprog.o

Details

Versions of the backend used inside the container:

• NVIDIA GPUs: CUDA 11.8 via NVIDIA base container
• AMD GPUs: ROCm 5.5.1
• OpenMP CPUs (used for AdaptiveCpp CPUs compilation and runtime): LLVM 17.0.0
• OpenCL Devices (used for DPC++ CPUs runtime): OpenCL via oneAPI DPC++ Get Started Guide

Tools:

• CMake 3.27
• vim 8.1
• git 2.25.1
• wget, unzip, python, nano...

Known issues

• clang++-17 (llvm) and clang++ (Intel llvm, the SYCL compiler) are not the same

• On some singularity configuration: clang++ cannot write temporary file during the compilation phase. You need to set the TMPDIR pointing to a directory inside the writable container

• On some singularity configuration you need to mount /sys /dev /proc inside the container (SINGUALRITY_BIND="/sys,/dev,/proc") to be able to see the GPUs devices