[FFmpeg-devel] [PATCH 2/3] lavd: add teletext quantizer

[Andrey Turkin]

2016-01-13 22:32 GMT+03:00 Michael Niedermayer <michael at niedermayer.cc>:

> to calculate the phase
>
> the area that has alternating 0/1 values can be correlated with a
> sin() and a cos(), if you consider the 2 resulting values as x and y
> coordinates the angle they form from the origin is the phase.
> (this can be worded simpler using complex numbers)
>
> when above is done its essential that a whole multiple of cycles is
> used the important part is that the used sin(x) and cos(x) vectors are
> orthogonal, that is sum sin(x) * cos(x) over the x values considered is
> 0 cutting them off randomls would break that
> (its also possible to use windowing instead of a exact multiple of cycles)
>
> consider that the phase is 0 so our signal input is exactly
> cos(x) (if we define that as 0°), the dot product of that with cos(x)
> is "1" and sin(x) is 0
> now if our input is shifted by 90deg that is it matches sin(x)
> then the dot product with a cos(x) vector is 0 and sin(x) is 1
> ...
> (also i for simpliity normalized the values, sum of cos(x)*cos(x) for
> x from 0 to n would of course be larger than 1 for a larger n)
>

That should work; I wonder about runtime cost though. sin/cos for a given
frequency can be precomputated; atan2 can be turned to table or something.
Still there'd be about 32 multiplications to get a phase (over 4 periods).
That's more than 25000 multiplications per second per teletext standard.