Scale, Translate and Rotate read inputs first pixel to determine other input area of interest, creating a rendering interdependency between inputs. For the full-frame system it would mean more memory usage as it wouldn't let to determine beforehand all areas operations must render. It would require a specific inputs rendering order and operations life (init/deinit) will be longer as we cannot just render all areas on first read.
It makes sense to read input first pixel if it's a single constant pixel value but if it has variable pixels next ones may cause a greater dependency area. Current implemention seems incorrect.
This patch adds a method for trying to read an input as a constant value that doesn't need rendering. If it's not constant, we can just assume whole input area is used for sake of correctness.