topic Computational cost of FFT with IPP in Intel® Integrated Performance Primitives

Computational cost of FFT with IPP

alzhykhar — Thu, 12 May 2005 20:05:07 GMT

Hi!

I wonder is there a definitive way (thumb rules :)) to tell how
computational time (Tc) for FFT using IPP depends on block size.
The Good Old Rule says

Tc ~ k*N*log(N)

for N-point FFT, where N is power of 2, and k being a
proportionality constant that depends on the particular FFT
implementation. Here one does believe that the time to compute
FFT is (proportional to) the number of operations (mul's + add's)
only.

I made some benchmarking to figure out how Intel FFT routine
ippsFFTFwd_RToPerm_32f() behaves in dependence on FFT order
and got some strange (wrong?) results. For FFT order in region
from 6 (block_size=64) up to 15 (block_size=32768) there is not
any dependency at all: the time is the same!

Any comments would be appreciated.

Thanks,
Albert

PS. My environment:
- Pentium 4 (2.40GHz, 512Mb RAM)
- Linux-2.6.11 (Mandrake-10.1)
- Intel IPP v4.1
- gcc-3.4.1

Re: Computational cost of FFT with IPP

Vladimir_Dudnik — Sat, 14 May 2005 00:10:11 GMT

Hi,

please look at this post, might be it is the info you are looking for

http://softwareforums.intel.com/ids/board/message?board.id=IPP&message.id=1598#M1598

Regards,

Vladimir

Re: Computational cost of FFT with IPP

borix — Sat, 14 May 2005 11:34:11 GMT

Hi Albert, what do you mean "the same"? total time or per butterfly? total time for different fft orders could be the same only in the case the function does nothing and returns with a bad status, please checkthe status codes, then let us know. Thank you