C5515 fft / ifft sample code

I've forwarded this to the software experts. Their feedback should be posted here.

BR
Tsvetolin Shulev

Hi Tsvetolin Shulev,

Any updates?

I used matlab to test fft/ifft function , the result data are the same with input data.
But i used TI DSP C5515 cfft/cifft , the result data (recBuf_ifft) are different with input data(i2s).
Please help check this issues.


========================================
matlab code
========================================

y = wavread('Ping\200Hz.wav');

s = fft(y);
f = abs(s);

y2 = ifft(s);
f2 = abs(fft(y2));

subplot(3,3,1), plot(s);
subplot(3,3,2), plot(f); % wav Spectrum
subplot(3,3,3), plot(y); % wav waveform
subplot(3,3,4), plot(y2); % ifft waveform
subplot(3,3,5), plot(f2); % ifft Spectrum



========================================
// c5515 code
========================================
static Uint16 recBuf[512];
static Uint16 recBuf_fft[512];
static Uint16 recBuf_ifft[512];

fft_test()
{
for(i=0;i<512;i+)
{
recBuf[i] = *i2sRxLeft; // get data from i2s
recBuf_fft[i] = recBuf[i]; // get data from i2s
}


cfft (&recBuf_fft[0], 512/2, NOSCALE); // fft
for(i = 0; i < 512; i++)
{
ABS(recBuf_fft[i]);
}
cifft (&recBuf_ifft[0], 512/2, NOSCALE); // ifft
for(i = 0; i < 512; i++)
{
ABS(recBuf_ifft[i]);
}
}

Any updates? Please help fft/ifft issue

Hi SMART

If I understand your question, your problem is that when you run FFT and then IFFT the results are not the same as the input vector.

Before I start looking into the code, here are two reasons why it happens-

The first and obvious one is the limited accuracy of 16-bit arithmetic. You "lose" bits of accuracy in FFT and IFFT. This depends on the size of the FFT (and the IFFT), so for 1024 FFT the error may be up to 10 bits.  MATLAB may convert everything to floating point and does 32-bit or even 64-bit computation.  (I am not sure how MATLAB) does it calculations)

The second issue is the square of N in the FFT formula.   TI code does not divide the FFT by the square of N.   Some people just multiply by N or divide by N.  I am not going to explain it here, but you can do search and find out (or you already know it)

So - tell us how the error looks like - is it like random error between two elements or you see a ratio (like N) between the elements of the input and the output of FFT/IFFT

best regards

Ran

Hi Ran,

Thanks for your kindly reply, i understand the issue.
For now , i can use TI cfft / cifft to convert codec data , but i don't know how to convert codec data by cfft/cifft from time scale to frequency scale.
1) Please advice convert time scale to frequency scale .
2) Please advice filter 2k Hz by cfft/cifft.


======================
DSP : C5515
codec : AIC3254 / 8k sample rate / 16 bit / MONO