Xilinx IPs for DFT, FFT, LTE_FFT

Started by kaz 1 year ago4 replieslatest reply 1 year ago88 views

Just started looking into Xilinx Fourier transform ip and found out that there are three versions:

DFT: apparently useful for any resolution

FFT: power of 2 resolution, i.e. subset of DFT

LTE_FFT: looks like subset of FFT as I see equal resource for same configuration

If I am right then may I ask why LTE_FFT is made a subset. Can't we just configure FFT for that. 



[ - ]
Reply by martinthompsonApril 24, 2018

According to LTE fourier transforms include one length which is non-power-of-2 (specifically a 1536 point DFT).  That specific transform will have been carefully factorised to optimise performance and/or size compared to a naive DFT of 1536 points.

[ - ]
Reply by kazApril 24, 2018

Thanks for that Martin but I noticed the same option for 1536 is available with fft ip setup.

Moreover I assume they could have done one ip for DFT and fft and lte configured as required.

I just was worried I may have missed something especially regarding matlab model for testing for example. 

[ - ]
Reply by martinthompsonApril 25, 2018

Ahh, I see what you mean.. In that case, it's probably just a hangover from when the FFT only did powers of 2?

[ - ]
Reply by oliviertApril 25, 2018

As says martinthompson The LTE FFT contains this 1536, which adds a radix-3 stage in the FFT. If you want to be able to change dynamically the FFT size you will have to include this radix-3 stage whenever you will use it or not.

The standard FFT IP is JUST for 2^N FFT length, the 1536 (3x512) is not possible in this IP. Some of my customers prefer to use a 2048 FFT size and then an interpolation to recompute the 1536 bins.