You can read the ADSP 21k Family Application Handbook for more detail on a lot of their examples and libraries. (I'm not sure if this is available in electronic format on the web site). Anyways, the last but one stage is computed separately to optimize the butterfly operations. This stage happens to have only 2 butterflies for each group. So they combine the 2 butterflies into a single loop to reduce overhead, thus gaining some extra speed. The last stage has only 1 butterfly per group and this is again done as a special case to gain speed (at the cost of writing a little extra code) Cheers Bhaskar --- In adsp@y..., Prashanth Bhushan <pb@s...> wrote: > Hi, > > I was going through the FFT code hosted by Analog Devices where they > have benchmarked a 128 point real FFT to take just 1160 cycles on ADSP > 21062 ! > > The implementation is split into 4 parts. Out of N stages first part > computes the first 2 stages . Second part computes for N-4 stages , > third part computes the last but one stage and the 4th part computes the > last stage. > > I needed to know why the N -1th stage is computed separately. > > Thanks in advance > Regds > Prashanth

> Date: Tue, 15 Jan 2002 18:12:09 +0530 > From: Prashanth Bhushan <> > > Hi, > > I was going through the FFT code hosted by Analog Devices where they > have benchmarked a 128 point real FFT to take just 1160 cycles on ADSP > 21062 ! > > The implementation is split into 4 parts. Out of N stages first part > computes the first 2 stages . Second part computes for N-4 stages , > third part computes the last but one stage and the 4th part computes the > last stage. Sounds like they use Meyer-Schwartz algorithm. And if this is the case, > I needed to know why the N -1th stage is computed separately. on this stage it is possible to compute 4 radix-2 butterflies in a single loop iteration, which is not possible on any other stages out of M, where N=2**M. Regards, Andrew > Thanks in advance > Regds > Prashanth >

Hi, I was going through the FFT code hosted by Analog Devices where they have benchmarked a 128 point real FFT to take just 1160 cycles on ADSP 21062 ! The implementation is split into 4 parts. Out of N stages first part computes the first 2 stages . Second part computes for N-4 stages , third part computes the last but one stage and the 4th part computes the last stage. I needed to know why the N -1th stage is computed separately. Thanks in advance Regds Prashanth