Yes, I did this by doing the shuffle from argb64 to bgra64 with SSE2 first and then manually shifting one bit left while keeping the maximum at 65535 for each value. This works, but it's also very... "hacky" and slow.
Maybe it would be easier to disable the Trillions of Colors and instead allow floating point? That would take care of the weirdness and heighten the precision too.
Yes, thought about that too. That's the way I am doing it for premiere. But as FFmpeg doesn't accept float as input I'd have to convert float to uint16_t first, ... and another color conversion to the target format. Also slow and not really nice.