actually this does not work, because polling requires:
in
andi
brne
which takes 4 clock cycles, that, on average, means a 2 clock offset,
while interrupts occur during a rjmp, which takes 2 clocks, which means,
on average, only 1 clock cycle of offset
I am not interested in higher or lower overhead: I am just interested in
minimizing its variance.
Being late of a few clock cycles, is not a problem: it just means that
the picture will be slightly shifted toward the right border; but,
starting the picture 1 or 2 clock cycles before (on one line), and 1 or
2 clock cycles after, on another line, means that it just wiggles sligthly
left and right
sufficiently handles the back/front porch without disrupting the sync
signal
todo improvement: compute a sensible amount of cycles to waste during
the porch
not the whole screen, but vertically locked
actually it has some issues, there is a single dot (at the
end/beginning of a cycle) that messes up the horizontal sync and moves
all the picture... must be fixed
