SYCL-docker

A simple self-contained container with SYCL
Contains:

• hipSYCL (latest nightly)
  • LLVM 13.0.1 (hipSYCL deps)
• DPC++ (release 2021-12)
• CUDA 11.4

Build

A pre built image can be found in DockerHub. To get it, type

docker pull luigicrisci/sycl_cuda:11.4

If you want to build the image by yourself, type

docker build -t <NAME> .

and docker will create an image called <NAME>.

The image requires ~22GB, so make sure to have enough space

Execute a container with GPU support

Prerequisites

CUDA

A detailed guide on how to enable CUDA support for containers can be found here

• Windows Users:
  • Only Windows 10 Inside preview and Windows 11 are supported
  • To run CUDA, you will need a working WSL 2 installation. If you already have a WSL 1 installation, you can upgrade it by typing wsl --set-version <Distro> 2
    Any linux distribution listed in the nvidia guide should work, but Ubuntu 20.04 is preferred. Note that nvidia driver on Windows already contains the one for WSL, so DO NOT install the Nvidia driver on WSL as conflicts will arise.
  • Install Docker Desktop on windows and DO NOT install docker on the WSL explicitly.
  • Turn on the Docker WSL2 backend in docker desktop under settings -> general -> use the WSL2 based engine. If your selected distro is also the default one, no other steps are needed. In the other case, you will have to enable the WSL integration under settings -> Resources -> WSL integration

Run the container

To run the container, type

docker run -it --gpus=all --name <CONTAINER_NAME> <IMAGE_NAME>

where

• -it creates an interactive shell on the container
• --gpus=all add all the gpu found in the system to the container

Select backend

hipSYCL and DPC++ have different headers files, so it is better to not enable them togheter
In the container you will find two files in your $HOME folder, manage_dpcpp.sh and manage_hipsycl.sh, which can install and uninstall the corresponding backend.

E.g. to install hipsycl

source manage_hipsycl.sh install

and to uninstall it

source manage_hipsycl.sh uninstall

Do NOT enable both backend togheter as it could brick the container (Not tested, feel free to make a try)

Building a SYCL app

Knowing your GPU capability

Each Nvidia GPU comes with a specific CUDA capabilities, which you need to specify when building a SYCL app using the CUDA backend.
While you can easly check the CUDA capabilities for your device by checking your architecture online, here you can find simple app that queries it for you.

HipSYCL

HipSYCL uses an environment variable HIPSYCL_TARGETS=BACKEND:CAPABILITEIS to select the device to build agains.
For example, to build an app for a Nvidia RTX 2070, with cuda capabilities 7.5

HIPSYCL_TARGETS=cuda:sm_75

Other backend are: omp for cpu code and hip:ARCH for amd gpu.

To compile an app, just use the hipSYCL compiler syclcc.

More info can be found here.

DPC++

To build an app using DPC++, you will have to use the clang compiler and select the appropriate triple. For example, to build a SYCL app for a Nvidia RTX 2070, with cuda capabilities 7.5

clang++ -fsycl -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend --offload_arch=sm_75 <FILENAME>

where

• -fsycl enables the SYCL implementation
• fsycl-targets=TRIPLE specifies the target hardware.

More info can be found here.

Cmake integration

Both DPC++ and hipSYCL can be used with CMake build system. A sample app that shows how to integrate them can be found here