topic One more note: Matlab uses in Intel® Integrated Performance Primitives

Insufficient FFT output accuracy

AlekseyS — Thu, 24 Oct 2013 11:21:29 GMT

Hello!

I'm calculating FFT of a complex signal using ippsFFTFwd_CToC_32fc function with order=10 (1024) and IPP_FFT_DIV_FWD_BY_N flag in Init. After applying my custom window (based on WinBlackmanStd) to the input data, one element of the output data in frequency domain gets exactly zero value (0, 0). As Matlab says, that element is the smallest element in the output array, its absolute value is near others':

Sp(30:35)
ans =
1.0e-005 *
0.0754 0.1696 0.0231 0.1748 0.4586 0.6651

This matlab output shows elements surrounding the problem 32nd element, which becomes zero in IPP's FFT output. The dynamic range of samples in frequency domain is from 0.8e+2 to 0.2e-6. I tried to use ippAlgHintAccurate flag, but it had no effect. Is seems like internal IPP's algorithms have insufficient accuracy while computing FFT.

If it is necessary, I can provide input and window samples.

Best Regards,
Aleksey

The IPP function is 32 bit

Thomas_Jensen1 — Thu, 24 Oct 2013 16:11:45 GMT

The IPP function is 32 bit float = "Single"; this having approx. 6.5 significant digits, which could be the problem.

Did you try to run your code using ippsFFTFwd_CToC_64fc? (if it exists, I didn't check).
With it, you could see where your digits are truncated.

yes, exactly - pls try the

Gennady_F_Intel — Fri, 25 Oct 2013 05:22:38 GMT

yes, exactly - pls try the 64fc version of this function.

Thanks a lot! Yes, 64fc

AlekseyS — Fri, 25 Oct 2013 07:48:46 GMT

Thanks a lot! Yes, 64fc version of FFT works properly. I overestimated float 32-bit function accuracy for my input signal with large dynamic range.

One more note: Matlab uses

Igor_A_Intel — Mon, 28 Oct 2013 14:48:19 GMT

One more note: Matlab uses IPP FFT internaly for performance - so results must be the same.

Regards, Igor.