[QUESTION] Why does composite only support three channels?

[caffeinism]

What is your question?
I’m not asking this because I’m having trouble with something, but I had a question while using it. Is there some kind of optimized functionality that works exclusively on 3-channel? It doesn’t seem like it would be too difficult to extend this to channels 1 and 4, so I was wondering why it only supports 3-channel.

[dlesage-nvidia]

Thanks for raising this. There is no CUDA hardware or optimization constraint that inherently requires three-channel inputs.
This is a historical API limitation: Composite was designed for 8-bit, three-channel foreground/background images and a single-channel mask.

The mask is applied independently to each color channel:

dst = round(bg + (fg - bg) * mask / 255)

When four output channels are requested, the operator blends the three color channels and sets the output alpha channel to 255; it does not perform alpha compositing.

Supporting 1 → 1 should be straightforward. For 4 → 4, we would first need to define whether the mask should blend all four channels independently or whether the input alpha channels should participate in true alpha compositing, since those have different semantics.

Therefore, this is feasible as a feature extension rather than being prevented by a three-channel-specific optimization. Could you clarify your use case and the expected behavior for four-channel inputs?