MSP430FR6043: real FFT/iFFT calculation using LEA

Part Number: MSP430FR5994

Hello, 

We are evaluating real FFT/iFFT with MSP430FR6043, and found an issue as below:

• Convert time domain signal A to frequency domain signal B using msp_fft_fixed_q15().
• Convert frequency domain signal B back to domain signal C using msp_ifft_fixed_q15().
• It’s found that A is totally different from C, and it is not caused by precision lost.

Can you help to look into this?

Thanks for your help.

Chaofan

The test code is shown as below.

the test vector provided is [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25,26, 27, 28, 29, 30, 31]

after fft and ifft, it's changed to [-11, -10, 12, -11, 6, -3, 14, -6, 15, -3, 14, -10, 22, -5, 21, 0, 27, 2, 12, 1, 24, -1, 20, 2, 25, 11, 22, 12, 20, 9, 13, 12]

#include "msp430.h"

#include <math.h>
#include <stdint.h>
#include <stdbool.h>
#include "DSPLib.h"

#define SAMPLES 32

/* Input signal and FFT result */
DSPLIB_DATA(input,MSP_ALIGN_FFT_Q15(SAMPLES))
_q15 input[SAMPLES*2];

void main(void)
{
 int16_t idx;
 msp_status status;
 msp_fft_q15_params fftParams;

 /* Disable WDT */
 WDTCTL = WDTPW + WDTHOLD;

 for (idx = 0; idx < SAMPLES; idx++)
 {
 input[idx] = idx;
 }

 fftParams.length = SAMPLES;
 fftParams.bitReverse = true;
 fftParams.twiddleTable = msp_cmplx_twiddle_table_32_q15;

 /* Perform real FFT with fixed scaling */
 status = msp_fft_fixed_q15(&fftParams, input);
 msp_checkStatus(status);

 /* Perform real iFFT with fixed scaling */
 status = msp_ifft_fixed_q15(&fftParams, input);
 msp_checkStatus(status);

 /* End of program. */
 __no_operation();
}

Hi Chaofan,

I'll need to consult our DSPLib expert and get back to you.

BR,

Leo

As I read this, your input is (a) amplitude roughly 0.0009, very close to loss of significance and (b) is sampled in an extremely small window (indistinguishable from DC). I suspect the ifft output is just noise.

Try generating an input vector using sin().

[Edit: Fixed Nyquist reference.]

Any update on this issue? The test vector I provide is just an example, it seems the ifft output is always different no matter what the test vector is.

What other test vectors did you try?

I would not expect an FFT+IFFT to be bit-exact due to rounding errors. (It should be repeatable, though.) In the test case you posted, I expect rounding errors would dominate.