Seamless multi-screen OpenGL rendering with heteregeneous multi-GPU configuration on Windows XP

Question

On Windows XP (64-bit) it seems to be impossible to render with OpenGL to two screens connected to different graphics cards with different GPUs (e.g. two NVIDIAs of different generations). What happens in this case is that rendering works only in one of the screens. On the other hand, with Direct3D it works without problem, rendering in both screens. Anyone knows why is this? Or more importantly: is there a way to render in both screens with OpenGL?

I have discovered that on Windows 7 rendering works on both screens even with GPUs of different brands (e.g. AMD and Intel). I think this may be because of its display model, which runs on top of a Direct3D compositer if I am not mistaken. This is just a suposition, I really don't know if it is the actual reason.

If Direct3D would be the solution, one idea would be to do all the rendering with OpenGL to a texture, and then somehow render this texture with Direct3D, suposing it isn't too slow.

Answer 1

What happens in Windows 7 is, that one GPU, or same type GPUs coupled, render the image to an offscreen buffer, which is then composited spanning the screens. However it is (yet) impossible to distribute rendering of a single context over GPUs of different making. That would require a standardized communication and synchronization infrastructure, which simply doesn't exsist. Neither OpenGL or Direct3D can do it.

What can be done is copying the rendering results into the onscreen framebuffers of several GPUs. Windows 7 and DirectX have support for this built in. Doing it with OpenGL is a bit more involved. Technically you render to an offscreen device context, usually a so called PBuffer. After finishing the rendering you copy the result using GDI functions to your window. This last copying step however is very slow, compared to the rest of OpenGL operation.

Answer 2

Both NVIDIA and AMD have ways of allowing you to choose which GPU to use. NVIDIA has WGL_NV_gpu_affinity and AMD has WGL_AMD_gpu_association. They both work rather differently, so you'll have to do different things on the different hardware to get the behavior you need.