Running something in between doubling and quadrupling can make things a lot more complex. If doubling and quadrupling can be done "uninterrupted", I can do it directly in OpenCL without having to involve D3D in between. If shaders need to be run in between, there's a lot of interop back and forth going on. That might explain the performance difference. In any case, as I said before, the plan for the "final" sharpening solution will be to run all sharpening on the fully upscaled video and not in between. But the algos need some improvements to make that work well for big upscaling factors.
|