Calculation time for 2D-FFT on DM6437

Anthony Gonin

Hi all,

I am new to DSP and I do some benchmark to evaluate the calculation time of 2D-FFT on EVMDM6437.

To make it I use the TMS320C64x DSPLib and I set up and down a GPIO before and after the call of DSP_fft16x16t function to get the calculation time on an oscilloscope. With CCS I use the clock functionality to get the number of cycles during the fft's calculation.

My results are : 3931 cycles and 36ms for 600MHz and 512 points.

The following formula given in documentation : (10 * nx/8 + 19) * ceil[log4(nx) −1] + (nx/8 + 2) * 7 + 28 + nx/8 says that I should have 3190 cycles for the DSP_fft16x16t() function. so it's not so far from my result.

My goal is to do an 2D-FFT (for image processing) on 512x512 grayscale image in a very short time.

I would like to know if I am doing evaluation in the right way and if my results are good. And also if someone can tell me how to improve the calculation time for FFT.

Thanks.

Anthony

over 9 years ago

0 Sivaraj Kuppuraj over 9 years ago

TI__Mastermind 35645 points

Hi Anthony,

Thanks for your post.

Please refer the C64x+ DSPLIB Benchmarks for your reference and check the test results as below:

http://software-dl.ti.com/sdoemb/sdoemb_public_sw/dsplib/latest//exports/DSPLib_c64xTest_Report.html

Have you seen this wiki note that comes with usage example:

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

Have you seen the example provided with the DSPLIB (http://focus.ti.com/docs/toolsw/folders/print/sprc265.html)

([INSTALL_DIR]\c64plus\dsplib_v210\example\fft_example)

Which FFT kernel are you using?

The common challenge you could probably see is that the input data wouldn't be correctly scaled while using the FFT kernel.

Note: It is important that the expected o/p should be represented by the fixed point data format selected.

Thanks & regards,

Sivaraj K

-------------------------------------------------------------------------------------------------------

Please click the Verify Answer button on this post if it answers your question.

-------------------------------------------------------------------------------------------------------

0 Anthony Gonin over 9 years ago in reply to Sivaraj Kuppuraj

Prodigy 30 points

Thanks for your answer.

I used the example provided with de DSPLib to evaluate the calculation time and I only used the 16x16 kernel.

I also try the example provided here : http://processors.wiki.ti.com/index.php/FFT/IFFT:_2D_Implementation , but the calculation time was very high.

Sivaraj Kuppuraj said:

Have you seen this wiki note that comes with usage example:

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

I'll read this note this afternoon.

Thanks

Anthony

0 Anthony Gonin over 9 years ago in reply to Anthony Gonin

Prodigy 30 points

Hi,

Since last time, I restart from scratch and I use the given example of the FFT.

To see if everything was OK I looked at output's signals and compared them to the given picture refOutput.jpg

Here is the first problem, as you can see on the following picture I haven't the same signals for 16x16 and 16x32.

Compare to the given picture :

If anyone can tell me why these output signals are different whereas they are calculated with the same input signal.

Thanks.

Anthony

Processors

Processors forum

Calculation time for 2D-FFT on DM6437