Other Parts Discussed in Thread: OMAPL138
Tool/software: Starterware
Hi,
The FFT function DSPF_sp_fftSPxSP in the DSPLIB produces wrong results. I am using the following hardware and software:
- TMDSLCDK6748
- CCS Version: 7.2.0.00013
- processor_sdk_rtos_omapl138_4_00_00_04 (RTOS 4)
- dsplib_c674x_3_4_0_0
I am using the provided FFT example: Benchmark_FFT_lcdkOMAPL138_c674ExampleProject that came with RTOS 4 and made the following changes:
/* ---------------------------------------------------------------- */
/* Initialize input vector temporarily. */
/* ---------------------------------------------------------------- */
for (i = 0; i < N; i++)
{
// x_cn[PAD + 2 * i] = sin (2 * 3.1415 * 50 * i / (double) N); //OLD
// x_cn[PAD + 2 * i + 1] = sin (2 * 3.1415 * 100 * i / (double) N); //OLD
x_cn[PAD + 2 * i] = sin(2.0 * PI * ((double)(i)) / (double)(8) ); //NEW
x_cn[PAD + 2 * i + 1] = 0; //NEW
#if defined(_TMS320C6600)
x_sa[PAD + 2 * i] = x_cn[PAD + 2 * i];
x_sa[PAD + 2 * i + 1] = x_cn[PAD + 2 * i + 1];
#endif
x_i [PAD + 2 * i] = x_cn[PAD + 2 * i];
x_i [PAD + 2 * i + 1] = x_cn[PAD + 2 * i + 1];
}
and
if (N <= 16) //NEW
{ //NEW
AUDIO_log("\n\nDSPF_sp_fftSPxSP, N=%d, radix=%d\n", //NEW
N, rad); //NEW
for(i = 0; i < (2 * N); i++) //NEW
{ //NEW
AUDIO_log("Input x_i[%3d]=%14.6f\n", //NEW
i, ptr_x_i[i]); //NEW
} //NEW
} //NEW
DSPF_sp_fftSPxSP (N, &ptr_x_i[0], &ptr_w[0], ptr_y_i, brev, rad, 0, N);
if (N <= 16) //NEW
{ //NEW
for(i = 0; i < (2 * N); i++) //NEW
{ //NEW
AUDIO_log("Output y_i[%3d]=%14.6f\n", //NEW
i, ptr_y_i[i]); //NEW
} //NEW
} //NEW
now the output of the program for N=8 is:
[C674X_0] /* ---------------------------------------------------------------- */ /* Single Precision FFT Benchmarks for N number of samples */ /* ---------------------------------------------------------------- */ DSPF_sp_fftSPxSP Iter#: 1 DSPF_sp_fftSPxSP, N=8, radix=2 Input x_i[ 0]= 0.000000 Input x_i[ 1]= 0.000000 Input x_i[ 2]= 0.707107 Input x_i[ 3]= 0.000000 Input x_i[ 4]= 1.000000 Input x_i[ 5]= 0.000000 Input x_i[ 6]= 0.707107 Input x_i[ 7]= 0.000000 Input x_i[ 8]= 0.000000 Input x_i[ 9]= 0.000000 Input x_i[ 10]= -0.707107 Input x_i[ 11]= 0.000000 Input x_i[ 12]= -1.000000 Input x_i[ 13]= 0.000000 Input x_i[ 14]= -0.707107 Input x_i[ 15]= 0.000000 Output y_i[ 0]= 0.000000 Output y_i[ 1]= 0.000000 Output y_i[ 2]= -2.000000 Output y_i[ 3]= -2.000000 Output y_i[ 4]= 0.000000 Output y_i[ 5]= -0.000000 Output y_i[ 6]= 2.000000 Output y_i[ 7]= 2.000000 Output y_i[ 8]= 0.000000 Output y_i[ 9]= 0.000000 Output y_i[ 10]= -2.000000 Output y_i[ 11]= 2.000000 Output y_i[ 12]= 0.000000 Output y_i[ 13]= -0.000000 Output y_i[ 14]= 2.000000 Output y_i[ 15]= -2.000000 Intrinsic Successful N = 8 radix = 2 natC: 310 optC: 199465 DSPF_sp_fftSPxSP Iter#: 2
The input is a real signal: a sine with frequency= f_sample / 8. The frequency is chosen in such a way that we should get exact results:at frequencies f_sample / 8 and 7 * f_sample / 8 we should get a value > 0 and all other frequencies should be 0. This means for N=8 and having a complex output array, the values with indices 2 & 3 and 14 & 15 should get a value > 0 and all other indices should be 0. But this is not the case: there are also values at indices 6 & 7 and 10 & 11 (frequencies 3 * f_sample / 8 and 5 * f_sample / 8).
What is wrong?
I realise that there are more efficient ways for an FFT on real signals, however, an FFT for complex signals should also work for real signals.
Remark: There is another problem and it occurs when the added print statements make the program slower. The following instructions need to be added to the beginning of main:
/*
* Clear CSR and IER register. This is necessary to prohibit interrupts
* from being enabled and left-over interrupts from being serviced.
*/
__asm("\t ZERO\t\t B0");
__asm("\t MVC\t\t B0, IER");
__asm("\t MVC\t\t B0, CSR");
For more details see post "TMDSLCDK6748: Program from new CCS project hangs".
Best regards,
Ad
