Am 18.09.2009 um 10:41 schrieb Chris Wood:
They also seem to need two output channels for each input, which
could be
more of an issue.
Uhm? Maybe you mean the real/imaginary parts? (Though, the input
has them,
too.)
Should be easy: we could abuse the color channels, since the
transform will
always be done separately for individual color components.
Yeah, the real/imaginary parts. The problem is that you do the
transform on each colour channel separately, so for rgb you need 6
channels (12 for the two buffers), which likely means 4 buffers in
practice.
Is this a problem? I'd split the RGB components and feed 3 identical
macro instances with them, each containing buffer and CI filters for
the FFT and the spectral processing and stuff, finally recombine the
components.
Concerning power-of-2: You can do things like convolution or cross-
correlation always on a power-of-2 basis. You need zero-padding anyway
to perform _noncyclic_ convolution or correlation: Keep in mind that
with DFT, you always get a spectrum of a periodic extension of your
signal. A bit more zero-padding than needed will not be wrong.
Another thing: Shouldn't we care about precision? Which GPU has how
many bits per color, and will it be sufficient for FFT?
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Quartzcomposer-dev mailing list ([email protected])
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/quartzcomposer-dev/archive%40mail-archive.com
This email sent to [email protected]
