• State
  • Locked Locked
  • Replies 5 replies
  • Subscribers 103 subscribers
  • Views 651 views
  • Users 0 members are here
Support feedback
Options
Options
Similar topics

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

DSPF_sp_fftSPXSP() multipass activation

Arye Lerner
Expert 1130 points

Hello,

I am trying to use the dsplib build in function (c674x dsplib_v10) of fft and inverse fft.

When I perform transform in a single pass implementation, it works, for array of 128 complex values, but when I try multipass implementation for 512 complex values in radix 4 format, I get QNaN values in  arrays after fft transform.

This is the code I use for multipass implementation:

Cmplx_input1() is an input array with real and imaginary values at even and odd indices, respectively. FFT_input1()  is output array. N_B  = 512 is a length of an array in complex values. a twiddle_factor() is calculated for full array of N_B values from function given in library. brev is an array also given in library.

    DSPF_sp_fftSPxSP (N_B, &Cmplx_input1[0],    &twiddle_factor[0],    FFT_input1, brev, N_B/2,  0,    N_B);

    DSPF_sp_fftSPxSP (N_B, &Cmplx_input1[2*3*N_B/4], &twiddle_factor[2*3*N_B/4], FFT_input1, brev, 4,   N_B*3/4, N_B);
    DSPF_sp_fftSPxSP (N_B, &Cmplx_input1[2*2*N_B/4], &twiddle_factor[2*3*N_B/4], FFT_input1, brev, 4,   N_B*2/4, N_B);
    DSPF_sp_fftSPxSP (N_B, &Cmplx_input1[2*1*N_B/4], &twiddle_factor[2*3*N_B/4], FFT_input1, brev, 4,   N_B*1/4, N_B);
    DSPF_sp_fftSPxSP (N_B, &Cmplx_input1[0],    &twiddle_factor[2*3*N_B/4], FFT_input1, brev, 4,   0,    N_B);

    DSPF_sp_ifftSPxSP (N_B, &Cmplx_input1[0],    &twiddle_factor[0],    FFT_input1, brev, N_B/2,  0,    N_B);

    DSPF_sp_ifftSPxSP (N_B, &FFT_input1[2*3*N_B/4], &twiddle_factor[2*3*N_B/4], Cmplx_input1, brev, 4,   N_B*3/4, N_B);
    DSPF_sp_ifftSPxSP (N_B, &FFT_input1[2*2*N_B/4], &twiddle_factor[2*3*N_B/4], Cmplx_input1, brev, 4,   N_B*2/4, N_B);
    DSPF_sp_ifftSPxSP (N_B, &FFT_input1[2*1*N_B/4], &twiddle_factor[2*3*N_B/4], Cmplx_input1, brev, 4,   N_B*1/4, N_B);
    DSPF_sp_ifftSPxSP (N_B, &FFT_input1[0],   &twiddle_factor[2*3*N_B/4], Cmplx_input1, brev, 4,   0,    N_B);

All the arrays is of size N_B*2

Please help

Thanks

Arye

  • 0
    Expert 1130 points

    Please disregard the code in first message, it was mistaken:

    // foward part:

        DSPF_sp_fftSPxSP (N_B,    &Cmplx_input1[0],                   &twiddle_factor[0],                  FFT_input1, brev, N_B/4,            0,    N_B);

        DSPF_sp_fftSPxSP (N_B/4, &Cmplx_input1[2*3*N_B/4], &twiddle_factor[2*3*N_B/4],  FFT_input1, brev,         4,   N_B*3/4, N_B);
        DSPF_sp_fftSPxSP (N_B/4, &Cmplx_input1[2*2*N_B/4], &twiddle_factor[2*3*N_B/4],  FFT_input1, brev,         4,   N_B*2/4, N_B);
        DSPF_sp_fftSPxSP (N_B/4, &Cmplx_input1[2*1*N_B/4], &twiddle_factor[2*3*N_B/4],  FFT_input1, brev,         4,   N_B*1/4, N_B);
        DSPF_sp_fftSPxSP (N_B/4, &Cmplx_input1[0],                  &twiddle_factor[2*3*N_B/4],  FFT_input1, brev,         4,             0,    N_B);

    // inverse part:

        DSPF_sp_ifftSPxSP (N_B,     &FFT_input1[0],                  &twiddle_factor[0],                 Cmplx_input2, brev,  N_B/4,               0,    N_B);

        DSPF_sp_ifftSPxSP (N_B/4, &FFT_input1[2*3*N_B/4], &twiddle_factor[2*3*N_B/4], Cmplx_input2, brev,          4,   N_B*3/4,   N_B);
        DSPF_sp_ifftSPxSP (N_B/4, &FFT_input1[2*2*N_B/4], &twiddle_factor[2*3*N_B/4], Cmplx_input2, brev,          4,   N_B*2/4,   N_B);
        DSPF_sp_ifftSPxSP (N_B/4, &FFT_input1[2*1*N_B/4], &twiddle_factor[2*3*N_B/4], Cmplx_input2, brev,          4,   N_B*1/4,   N_B);
        DSPF_sp_ifftSPxSP (N_B/4, &FFT_input1[0],                  &twiddle_factor[2*3*N_B/4], Cmplx_input2, brev,          4,                0,   N_B);

    It does not gives Qnan now, but the conversion still is totaly wrong.

    This code is the same as in SPRA947a, wheere the example is given on fft only. I assumed that inverse fft will be of the same format... or not?

     

    Thanks

    Arye

  • 0 in reply to Arye Lerner
    Prodigy 30 points

    The same problem guy, it looks does not work

    It pass fthe irst time but any next time it returns the same value(***)

  • 0 in reply to Milun
    Expert 1130 points

    Hi,

    I sort of figured that out.

    You need to define some of the arrays you use as double word alligned. If I understand it right, it uses space in between for calculations. That can explain the data being written back to the array if not alligned.

    Anyway, just define the arrays that need to be alligned (look at the function description) as:

    #pragma DATA_ALIGN(buffer, 128)      // or 64 instead of 128

    float buffer[SIZE];

     

    Regards

    Arye

  • 0 in reply to Arye Lerner
    Prodigy 40 points
    Recently I met the same problem in applying DSPF_sp_fftSPXSP() multipass activation as Arye did. whatever methods I tried, the output of DSPF_sp_fftSPXSP() with multipass activation is not the same as that of DSPF_sp_fftSPXSP() with single-pass implementation, which is correct compared with the result using MATLAB simulation. I really need your help and please help me figure it out. Thanks a million
  • 0 in reply to James Wang
    TI__Mastermind 35645 points
    Hi James,

    Please follow-up this issue on your new thread created as below:

    http://e2e.ti.com/support/dsp/omap_applications_processors/f/42/t/389618

    Thanks & regards,
    Sivaraj K