SPRACN4 August   2019 66AK2G12 , 66AK2H06 , 66AK2H12 , 66AK2H14 , OMAP-L132 , OMAP-L138 , TMS320C6452 , TMS320C6454 , TMS320C6455 , TMS320C6457 , TMS320C6652 , TMS320C6654 , TMS320C6655 , TMS320C6657 , TMS320C6672 , TMS320C6674 , TMS320C6678 , TMS320C6742 , TMS320C6743 , TMS320C6745 , TMS320C6746 , TMS320C6747 , TMS320C6748

 

  1.   Using DSPLIB FFT Implementation for Real Input and Without Data Scaling
    1.     Trademarks
    2. 1 Real Input Introduction
      1. 1.1 Prerequisites
      2. 1.2 Computing a Length N/2 Complex FFT From a Length N Real Input Sequence
      3. 1.3 Returning to a Length N Real Sequence Using a Length N/2 Complex IFFT
      4. 1.4 Benchmark of the Efficient Compute of FFT
    3. 2 Fixed Point FFT With No Data Scaling
      1. 2.1 Suggested Change
      2. 2.2 Example Application
    4. 3 Summary
    5. 4 References

3 Summary

  • N = 256. Since 256 = 4^4, the processing will happen as follows: 3x radix 4 iterations in the main loop (scaling by 2) and 1x radix 4 iteration at the end. So the total scaling will be 2^3. This is because every stage applies scaling of 2 but the last stage doesn’t apply any scaling. Thus, scaling by 2*2*2 = 8 is expected. The images confirm the understanding. Note the X Output magnitude increased by 8 times when scaling is removed from FFT and then from iFFT.
  • When both the kernels are updated to not apply scaling, the X output magnitude is 256 times the input. This is correct as 1/N scaling is not applied overall. Thus, X Output to be 256 times X Input is expected.

Thus, the suggested changes correctly remove the scaling from the two kernels.