I think Chris' FFT is down to about 3% of the total cost. So even if
we found an FFT that was twice as fast (very unlikely), it would only
save 1.5%. Thus I dont think linking to a big package like FFTW is
worth the trouble.
Just to clarify some other points:
layer3 works with 1152 sample 'frames', but all the analysis is
done on two 576 sample 'granules'. Except for the frame header
information, the two granules in each frame are completely independent.
For each 576 sample granule, the psy-model does one 1024 FFT centered
over the granule, and 3 256FFT's centered over each third of the
granule. So I dont know if 1024 is really optimal, maybe 768 would
have been a better choise.
Mark
--
MP3 ENCODER mailing list
