I configure the SPU to do 256-point range FFT with input data from RIF, and then compare FFT result between SPU and other FFT library (Python numpy.ff...
Show More
I configure the SPU to do 256-point range FFT with input data from RIF, and then compare FFT result between SPU and other FFT library (Python numpy.fft).
For some input data as in picture below, upper plot is time domain signal of small noise fluctuation, ADC value varies from 1750 to 2000. The lower plot is FFT result from SPU and numpy.fft, and they aligns with each other so well that we can only see one green line. (except for one point at bin 0).
This shows the whole setup and SPU configuration is somewhat correct.
However, for input data with larger fluctuation as in following picture (noise + 100kHz sine wave with ADC value varies from 1920 to 2120), the FFT difference between SPU (red line) and numpy.fft (green line) is huge. SPU result shows much higher noise floor.
Moreover, if I further increase the input amplitude as in following picture (ADC value varies from 1400 to 2600), surprisingly SPU results' noise floor reduced, but 100kHz frequency's amplitude also reduced, and some artifact frequency component showed up.
Here is my SPU configuration:
/* INPUT SOURCE */
cfg->input.dataSource = IfxSpu_DataSource_rif0;
cfg->input.rif.numAntennae = IfxSpu_Num_Antennae_2;
cfg->input.rif.dataFormat = IfxSpu_InputDataFormat_real;
cfg->input.rif.dataType = IfxSpu_InputDataType_unsigned;
cfg->input.rif.sampleCount = 256;
cfg->input.rif.numRamps = 256;
IfxSpu_setupInput(slot, &cfg->input);
/* BIN-REJECTION */
cfg->binrej.mode = IfxSpu_BinRejection_Mode_reject;
cfg->binrej.numAllowedBins = 128;
cfg->binrej.thresholdEnabled = FALSE;
cfg->binrej.thresholdValue = 0xFFFFu;
IfxSpu_setupBinRejection(slot, &cfg->binrej);
IfxSpu_PassId passId = IfxSpu_PassId_0;
IfxSpu_PassConfig *pcfg = &cfg->pass[passId];
/* MATH 1 */
pcfg->math1.loaderExponent = 5;
pcfg->math1.numDropFirstSamples = 0;
pcfg->math1.numDropLastSamples = 0;
pcfg->math1.numPadFrontSamples = 0;
pcfg->math1.window.enabled = FALSE;
pcfg->math1.window.dataFormat = IfxSpu_WindowDataFormat_real16;
pcfg->math1.window.baseAddress = IFX_OFFSETOF(SPU_Cmem0_Map_t, rWndw);
for (antNr = 0; antNr < RADAR_NUM_RX; antNr++)
{
pcfg->math1.window.antennaOffsets[antNr] = 0;
}
pcfg->math1.phaseShift = IfxSpu_PhaseShift_0;
IfxSpu_setupMath1(slot, passId, &pcfg->math1);
/* FFT ENGINE UNLOADER */
pcfg->fft.enabled = TRUE;
pcfg->fft.inversed = FALSE;
pcfg->fft.size = IfxSpu_getSizeCode(RADAR_FFT1_LEN);
pcfg->fft.dataFormat = IfxSpu_FftDataFormat_complex32Bit;
pcfg->fft.exponent = 0;
pcfg->fft.forceToReal = FALSE;
IfxSpu_setupFft(slot, passId, &pcfg->fft);
/* FFT OUTPUT */
pcfg->fftOut.enabled = TRUE;
pcfg->fftOut.baseAddress = IFX_OFFSETOF(SPU_Emem_Map_t, fft_range);
pcfg->fftOut.format = IfxSpu_ODP_Format_complexHalfFloat;
pcfg->fftOut.exponent = 16;
pcfg->fftOut.inPlace = FALSE;
IfxSpu_setupFftOutput(slot, passId, &pcfg->fftOut);
I tried changing math1.loaderExponent from 0 to 18, it only changes the scale of FFT result, and do not solve the problem.
I also tried changing fftOut.exponent, but it does not have any effect.
Please help.
Thank you very much.
Show Less
Solved
