Developing code which can run on AMD and NVIDIA GPUs

Discussion Zone
, ,
1

The clusters I have access to have a mix of AMD and NVIDIA GPUs. How can I develop code which is easy to run with any accelerator?

2

One option is to use ILGPU. This is a project built on top of C# which makes it easy to target code to both AMD and NVIDIA GPUs.

Start by installing dotnet (C# shares some of the same tools with F# and other Microsoft languages).
If you are on a cluster you will want to install via bash script which does not require admin privileges.

Otherwise you can install via:

After install, you should be able to run dotnet --info from your command line.

ILGPU has some documentation for more details but let’s run a simple program here to get started.

Start by running git clone https://github.com/m4rs-mt/ILGPU to get the ILGPU source code.

Now, cd ILGPU/Samples/DeviceInfo.
Now edit the DeviceInfo.csproj file so that
The line beginning: <TargetFrameworks> reads:

<TargetFrameworks>net5.0</TargetFrameworks>

To build this project start by restoring it:

dotnet restore

Then build it & run it

dotnet build -c Release
dotnet run -c Release --framework net5.0

You have now run your first sample of ILGPU + C#. Try running some of the other samples on computes with AMD or NVIDIA GPUs.