ADI's benchmarks, and comments made in this forum, indicate cfftf should take 11071 cycles for a 1024-point FFT, assuming the real data and the imaginary data are in separate memory banks, and the FFT in/out array is 1024-word aligned. But my testing on a 450 MHz 21469-based platform indicates cfftf takes about 29000 cycles (based on checking the EMUCLK register before and after).
Other than audio DMAs (4 channels, 48000 Hz) there is very little else that would be using memory bandwidth. My executable code is in internal memory.
Are there any other restrictions (eg. caching, or in which block the code resides) that might affect the timing by such a large factor?
Moving to CrossCore Embedded Studio and Add-ins
Further measurements indicate that although the average number of cycles is high, the minimum is very close to the 11071 cycles I expected. My guess is that ISRs are occasionally getting in there and inflating the cycle count. You can probably close this; if it turns out that ISRs are not the problem I'll open another with further details. Thx