Is it possible create torch.Tensor from IDirect3DSurface or CUDA memory directly?

[GF-Huang]

In details, I use Windows Graphics Capture technology to capture specific app window frames. This is equivalent to screen recording, but only for a specific window. You can read this repo if you interest it.

At currently, I copy the frame bitmap(raw RGBA byte array) from GPU to CPU, then create tensor by torch.tensor(rawRgbaByteArray, ...) (in other words, copy from CPU to GPU). Obviously this is unnecessary.

So, is it possible create torch.Tensor from IDirect3DSurface directly? That is copy from GPU buffer to another GPU buffer by VRAM copying rather than RAM copying, if I understand correctly.

[dayo05]

I don't have enough knowledge of this but I found some reference might related to this:

Convert Cuda memory to LibTorch tensor
Create TorchSharp tensor from raw pointer

If you mix those with P/Invoke, it seems able to create Tensor without coping entire data to CPU.