DSPLIB FFT rounding problem

ayhan_kucukmanisa

Hi. im using dm6437 board. i'm using dsplib DSP_fft16x16 fft function. i compared matlab fft result and dsp fft result. results are like these :

matlab fft : 496.00 , -16.00 + 162.00i , -16.00 + 80.00i , -16.00 + 53.00i , -16.00 + 39.00i , -16.00 + 30.00i

dsp fft : (496.000000,0.000000),(-16.000000,160.000000),(-16.000000,80.000000),(-16.000000,52.000000),(-16.000000,40.000000),(-16.000000,24.000000)

i understand DSP_fft16x16 function is making rounding it's OK .but i think this rounding that in somewhere is big. is there anything that i have to do for this or i miss something.

Or fft function result must be like that?

Thanks!

over 14 years ago

0 Gagan Maur over 14 years ago

TI__Expert 8150 points

Ayhan, can you confirm if the issue is resolved or you still have the issue?

Gagan

0 ayhan_kucukmanisa over 14 years ago in reply to Gagan Maur

Intellectual 560 points

i couldnt solve rounding problem gagan. and i wanna add something. between matlab fft result and dsp fft result there is a scale difference. for example. dsp fft result for 16 points 60 matlab fft result is 120, and for 32 point dsp result is 80 matlab fft result is 320 ext.

how can i solve this?

0 Gagan Maur over 14 years ago in reply to ayhan_kucukmanisa

TI__Expert 8150 points

For removing scaling in the code, please see here:

http://processors.wiki.ti.com/index.php/FFT_Implementation_With_No_Data_Scaling

You may also choose to use 16x32 or 32x32 FFT.

Note, 16x16 fft is fixed point implementation and there will be differences b/w the floating point FFT that Matlab implements

Regards,

Gagan

0 Tiny over 14 years ago in reply to Gagan Maur

Prodigy 100 points

Dear Gagan:

Hi!

Now I use dsplib in my davinci DM6446 project. I have some questions.

1.16x16 fft How can I get the correct output? The output is short t precision or r Q.15 format. How can I understand it?

2.I input a 256*256 's picture (GRAY every pix has a 8-bits) . I input the 256*256 picture in the real components and the imaginary components are all 0. And I watch the output .The output : the real components and the imaginary components all have 16-bits . How can I get the correct results? How can I transform it correctly?

Regards,

Tiny

Processors

Processors forum

DSPLIB FFT rounding problem