Hi, list, I've completed a collection of objects using wavelet transforms for 2 uses: de-noising a signal, and subband adaptive filtering. So far, I've got the routines coded compiled, linked, and debugged, but they aren't really up to snuff for computations. I can't be sure if the adaptation algorithm is *really* working or not, since I'm limited in the size of blocks for convolution. I need to have a blocksize that is comparable to the room reverberation time, and at present the best I can do is about .4 seconds (at 22050 Hz). At the moment, the program runs with a blocksize of 16384 and overlap of 2 and sample rate of 22050, without any trouble, and for anything greater than that, introduces cuts in audio.
The big trouble at the moment is my convolution routines. I coded a naive convolution just to get it out of the way, so I could work on the other bits. At the moment, I'm trying to figure out fftw for use in a more efficient convolution. I was wondering if anyone uses fftw for Pd, and if they could offer a nuanced explanation of how to use the data types and fftw wisdom to get fast computations.
(I run FC4 with a 2GHz AMD 2800 Sempron (which I don't think uses SIMD, it uses x86/x87), and I'm using Pd-extended v38.4)
-- Charles Zachary Henry
anti.dazed.med Med student who needs a Mickey's