On Mon, May 28, 2007 at 03:15:33PM +0900, David Cournapeau wrote: > Matthieu Brucher wrote: > > > > There really aren't any transparent fast fft convolutions in > > SciPy. The closest thing is in signaltools, fftconvolve, and if > > you ask it to convolve, say, sequences whose length add up to > > 7902, then it will do a size 7901 transform. > > > > > > BTW, is this really a glitch ? I think there is two schools there : > > - those who think the software must do something > > - those who think it is the programmer's responsibility. > > > > If I give to a fftconvolve sequences that add up this way, I want it > > to use a 7901 transform, not a 2^n transform. So you understood I'm in > > the second group. > I may miss something, but in theory at least, using zeros padding and > fft of size 2^N (8192 here) will give you exactly the same result that > doing fft of size 7901 if done right. Then, there are some precision > problems which may appear, but I don't think they are significant for > common cases. For example, I know for sure that matlab implements fast > convolution/correlation this way (eg in lpc code, for example).
There is also the problem that you suddenly double your memory usage. Stéfan _______________________________________________ Numpy-discussion mailing list [email protected] http://projects.scipy.org/mailman/listinfo/numpy-discussion
