And an update/clarification in case anyone ever finds this thread: I missed the fact that gfxanalyzer incorporates the 1/fftsize in the windowing function, as you apparently need to do because when using a non-unity ("rectangular") window you don't scale by 1/fftsize but a different value dependent on the window (one over the sum of the window values.) You apply the scaling (or the scaling+window) on the sample values before the fft. Then when you do the fft you can calculate the magnitude (in absolute scale, 0 to 1) as shown above.
Interestingly, though, you don't need to scale back up when coming the other way. Meaning, you put your sample values in the fft workspace array (paired with 0-value complex values) at 1/fftsize scale, do the fft, do fft_permute, do whatever you want to to the data, then fft_ipermute, then ifft, and the resulting real values in the workspace do not need to be scaled up by fftsize, they can just be used as-is. This quirk led to a lot of confusion on my part.
Someone correct me if I've missed something else. :-)
|