Updating a texture used in a post-process with data directly from the GPU without a CPU copy.

Development Rendering
, , , , , ,
1

Hello everyone, I hope this is the right place to ask this question,

I am currently trying to include the Nvidia OptiX framework via a Plugin into Unreal, but I’m a little stuck on executing a performant texture update each frame.

What I am currently (successfully) doing is this:

  1. OptiX uses the Unreal camera matrices, traces and renders its output each frame into an full-screen sized (OptiX) buffer on the GPU.

  2. Then, I map this output buffer, which consists of FColor values (or float depth values for an additional depth buffer) and update the Unreal Texture2D (in the render thread) via:

    RHICmdList.UpdateTexture2D(OutputTextureColorRef, 0, TextureRegion, Width * 4, (uint8*)OptiXData);

  3. The respective textures are used in a post process material and blended over the standard scene each frame.

All of this works as expected. What surprised me now is that the OptiX full screen trace takes almost less time than the mapping and updating part, especially when using a VR headset. 
I am assuming that this is because the data gets copied to the CPU and then straight back into the texture buffer on the GPU again when updating the texture.

Now, my question is: Can this copy somehow be avoided? I see two possibilities but am completely lost on how viable they actually are:

  1. Use the Unreal Texture2D and somehow get a native pointer to the memory on the GPU, and tell OptiX to write its output into this memory. Is the former part even possible here?
  2. Copy the data in the OptiX buffer directly into the memory of the Texture2D without going through a CPU copy. OptiX has no direct interoperability with DirectX textures, but CUDA can be used as a bridge here. Is it possible to directly update the underlying resource of an Unreal Texture2D here?

My knowledge about GPU Memory Management and how Unreal does it is sadly very limited - I would be really glad if someone here could help me out. 
Alternatively I thought about just using the new DXR functionality directly, but the current implementation in Unreal seems to be limited still, especially considering VR devices.

Thanks in advance, 
David