The GPU device instance has been suspended

Hi, I am playing around with your library and wanted to compare performance of two of my GPUs.

For this I am doing the following:

List<GraphicsDevice> gpus = Gpu.EnumerateDevices().ToList();

for (int g = 0; g < gpus.Count; g++)
{
    Stopwatch stopwatch = new Stopwatch();
    
    stopwatch.Start();
    this.DoComputation(gpus[g]);
    stopwatch.Stop();

    Console.WriteLine($"GPU {g}: {stopwatch.Elapsed}.");
}

But I am getting the following error:

Unhandled exception. SharpGen.Runtime.SharpGenException: HRESULT: [0x887A0005], Module: [Vortice.DXGI], ApiCode: [DXGI_ERROR_DEVICE_REMOVED/DeviceRemoved], Message: The GPU device instance has been suspended. Use GetDeviceRemovedReason to determine the appropriate action.

   at SharpGen.Runtime.Result.CheckError()
   at Vortice.Direct3D12.ID3D12Device.CreateCommittedResource(HeapProperties& heapPropertiesRef, HeapFlags heapFlags, ResourceDescription& descRef, ResourceStates initialResourceState, Nullable`1 optimizedClearValueRef, Guid riidResource)
   at Vortice.Direct3D12.ID3D12Device.CreateCommittedResource(HeapProperties heapProperties, HeapFlags heapFlags, ResourceDescription description, ResourceStates initialResourceState, Nullable`1 optimizedClearValue)
   at ComputeSharp.Graphics.Buffers.Buffer`1..ctor(GraphicsDevice device, Int32 size, Int32 sizeInBytes, BufferType bufferType)
   at ComputeSharp.Graphics.Buffers.Abstract.HlslStructuredBuffer`1.GetData(Span`1 span, Int32 offset, Int32 count)
   at ComputeSharp.Graphics.Buffers.Abstract.HlslBuffer`1.GetData(Int32 offset, Int32 count)
   at ComputeSharp.Graphics.Buffers.Abstract.HlslBuffer`1.GetData()

I know it’s an issue in the DirectX, but maybe you know how to solve it?

The same code works with Gpu.Default.

Also, is it possible to retrieve any info about the GPU? Like name, vRAM size, cores count?

---

My setup:

• Visual Studio 2019.
• .NET Core 3.0.0-preview8-28405-07.
• GTX 1080.
• GTX 1050 ti.
• Windows 10 Pro x64.

[Sergio0694]

Yeah you can just run the shader with 1 iteration, that works too 👍 The important thing is just that the number of iteration axes remains the same during your preloading and actual iteration. That’s the same reason why I haven’t added an API to preload a shader, as a shader also contains the number of threads to run along each axis (the ThreadNum HLSL attribute). Basically, you can do this:

• If your shader will be run just on the X axis, preload it with 1
• If it’ll run on X and Y, preload it with 2, 2
• If it’ll run on X, Y and Z, preload it with 2, 2, 2

As for your results, glad to hear you’re having such an improvement in your scenario! 🚀 If you want to squeeze out even more performance, you can also try to tweak the shader, eg. to process more than one output value in each iteration, so that the relative overhead for each value would be reduced. You’ll need to make some tests for this though, as results may vary depending on your specific input data, buffer size, etc. - but it might be worth a try if you want the best possible performance.

[Sergio0694]

No problem, happy to help! Glad we could get this fixed so quickly 😊