The code creates a frame with the timing information, and keeps track of the
times the applications does various things during the frame loop. Like
starting to render and finished rendering and so on. It keeps track of that
frame until it has determined that the application or the system compositor
will never do anything thing with that frame.
Now the allocation side of thing is very simple, just a long array of
FRAME_COUNT size that is reused with frame_id % FRAME_COUNT. So if the
application or the compositor keeps a frame along for too long it will try to
reuse that frame. The code has asserts to catch this. Making FRAME_COUNT makes
that re-use more unlikely.
If the last returned display time shifts backwards slightly with respect to the
last sampled display time from the compositor, the next predicted display time
will not move forward by one frame. Adding half the display period to the
comparison makes the pacing robust to this case.
