[API Proposal]: provide igp+npu access from dotnet

[mprevot]

Background and motivation

I am not sure if this goes do to dotnet roslyn or ruyJIT or both.

At the loud era of AI, it will be very interesting to be able to access to local igp (integrated GPU) and NPU accelerators (for instance) from dotnet code.
The API proposal will likely require significant work. I nevertheless suggest a few examples.

API Proposal

namespace System.Collections.Generic;

public abstract class GPUarray<T> : IEnumerable<T>
{
public abstract void Upload();
public abstract void Download();
public abstract bool IsOnDevice {get;}
public abstract GPUarray<T> Operation(Func<T,T>);
//etc
}

API Usage

BitmapImage b = Factory.GetImage();
b.ToGPUArray().Upload();
b.Operation(someFilterOnImage);
b.Operation(addAnotherImage);

//display
b.Show(UIElementTarget); //do not download, but inject in visual element

// or save b
b.Download();
b.Save(heifTargetFileWithParameters);

Alternative Designs

No response

Risks

No response

[colejohnson66]

A generic “operation” is way too abstract to mean anything. Not to mention that CPU code can not be executed on a GPU.

[jkotas]

Check the discussion in #8220. Number of independent projects provide GPU access from .NET. We have not found strong reason for these capabilities to be included in the core .NET runtime project.