I'm not sure that MaxCoeffNum works even in theory. If you cannot resolve the coefficients to the scale on which they vary between FFT bins, it seems to me you fundamentally cannot tell the difference between 2 frequencies that are separated using this coefficient. Although there may be other tricks to save resources (like only storing the LSBs of fine-grain coefficients), I don't believe that just capping the number of coefficients can work. Unless there is evidence to the contrary, I'd suggest removing the MaxCoeffNum functionality.
On Thu, Apr 30, 2009 at 11:52 AM, G Jones <[email protected]> wrote: > Hello, > I am finding that FFTs with sizes <= 2^11 work fine, but >2^11 are producing > invalid output. Roughly, a CW complex tone is appearing at two frequencies > instead of just at one. I can reproduce this in simulation and in the lab. > Andrew suggested this may be due to the MaxCoeffNum, and I now remember I > ran into this problem before but never got around to solving it. I > understand the idea behind MaxCoeffNum, that for large FFTs the differences > between twiddle factors become small enough that they can be neglected. > However, as it is currently implemented, it does not seem to work. I am > using the green 7.1 blocks. Has anyone successfully used this feature (i.e. > made a single input biplex FFT of length greater than MaxCoeffNum, 2^11 by > default)? > Thanks, > Glenn > -- Aaron Parsons 510-406-4322 (cell) 787-878-2612 x329 (arecibo office) 787-721-3991 (home)
