topic ipps fft function in Intel® Integrated Performance Primitives

ipps fft function

sever70 — Tue, 29 May 2012 07:39:55 GMT

Hi,

I am trying to take FFT of a real signal and also the output should be a real signal. Thus, I useone of these functions :

ippsFFTFwd_RToPerm_32f

ippsFFTFwd_RToPack_32f

ippsFFTFwd_RToCCS_32f

However the output signal doesn't match with the Matlab output of fft function. Any idea?

Thanks in advance

ipps fft function

Thomas_Jensen1 — Tue, 29 May 2012 08:56:39 GMT

What is it that doesn't match?

Numbers?

ipps fft function

sever70 — Tue, 29 May 2012 09:14:41 GMT

Yes, the numbers doesn't match..

ipps fft function

sever70 — Tue, 29 May 2012 09:21:07 GMT

Matlab output for a length of 8 array

4.000000

-1.414214

4.000000

1.414214

-4.000000

1.414214

4.000000

-1.414214

And by using these ippsfunctions,

ippsFFTInitAlloc_R_32f( &pSpec, 3, IPP_FFT_DIV_INV_BY_N, ippAlgHintNone );

ippsFFTFwd_RToCCS_32f(pX, pY, pSpec, pBuf);

I got these outputs:

4

-1.41421

-4

1.41421

-1.41421

ipps fft function

Thomas_Jensen1 — Tue, 29 May 2012 09:27:45 GMT

You should have taken a look at the IPP documentation. For CCS, your output is as expected.
Look at the IPP definition of CCP, you'll see it there:

http://software.intel.com/sites/products/documentation/hpc/compilerpro/en-us/cpp/lin/ipp/ipps/ipps_ch7/ch7_packed_formats.html#CCS

ipps fft function

sever70 — Tue, 29 May 2012 09:30:59 GMT

Thank you Thomas. But shouldn't there be a one to one match with matlab fft functions. Or which function I should use for this purpose?

Ipps FFT function

SergeyKostrov — Tue, 29 May 2012 12:27:04 GMT

Quoting sever70

Hi,

I am trying to take FFT of a real signal and also the output should be a real signal. Thus, I useone of these functions :

ippsFFTFwd_RToPerm_32f
ippsFFTFwd_RToPack_32f
ippsFFTFwd_RToCCS_32f
However the output signal doesn't match with the Matlab output of fft function. Any idea?

As far as I know the Matlab does calculations usingDouble-Precision variables. The set of IPPSfunctions
you're using are Single-Precision ( ..._32f ). Could you submit a complete IPP Test-Case?

Best regards,
Sergey

ipps fft function

Thomas_Jensen1 — Tue, 29 May 2012 15:24:48 GMT

There would only be a match if the output of matlap and of ipp was in the same "domain", I mean, of the same type.

FFT and DFT uses different data domains (types) for efficiency purposes.
For instance, if you data is purely real, you can hold its complex version is a so-called packed domain. This is much smaller the the full complex version and thus faster (better cache utilization etc.).
However, if you print packed data, it looks not at all like normal complex data. You would have to convert packed to complex before printing it.

I don't really know useful purpose of the CCS format, but if you simply select normal complex output format, and then print that (possible first converting to real format), I'm sure your comparison would understandable.

Do you understand my text?

ipps fft function

Ying_H_Intel — Wed, 30 May 2012 08:45:34 GMT

Hi Sever70,

Not sure Matlab's out put format. But same suggestion as Thomas,
The format is as below

Arrangement of Forward Fourier Transform Results in Packed Formats - Even Length
Index	0	1	2	3	. . .	N-2	N-1	N	N+1
Pack	R₀	R₁	I₁	R₂	. . .	I_(N-1)/2	R_N/2
Perm	R₀	R_N/2	R₁	I₁	. . .	R_N/2-1	I_N/2-1
CCS	R₀	0	R₁	I₁	. . .	R_N/2-1	I_N/2-1	R_N/2	0

Forward Fourier transform Result Representation in Packed Formats - Odd Length
Index	0	1	2	3	. . .	N-2	N-1	N
Pack	R₀	R₁	I₁	R₂	. . .	R_(N-1)/2	I_(N-1)/2
Perm	R₀	R₁	I₁	R₂	. . .	R_(N-1)/2	I_(N-1)/2
CCS	R₀	0	R₁	I₁	. . .	I_(N-1)/2-1	R_(N-1)/2	I_(N-1)/2

and you may use the Transform Functions to get real complex result.

Intel IPP Transform Functions
Function Base Name	Operation
Support Functions
ConjPack	Converts the data inPack format to complex data format.
ConjPerm	Converts the data in Perm format to complex data format.
ConjCcs	Converts the data in CCS format to complex

Best Regards,
Ying

ipps fft function

sever70 — Wed, 30 May 2012 12:34:17 GMT

Thanks Thomas and Ying.Finally I got correctFFT values for real data by usingCCS format.