Iterative Radix2-FFT on FPGA

Hey,

so now I'm ready with my first FFT Implementation.
Like I mentioned in my earlier Post I want to implement a Iterative Radix2 FFT Algorithm.
You can see a Pseudocode on Wikipedia: https://en.wikipedia.org/wiki/Cooley%E2%80%93Tukey_FFT_algorithm#Data_reordering,_bit_reversal,_and_in-place_algorithms

I compared it with a C++ Implementation of it and it's relative exact.
With relative I mean that there are rounding Errors in it or Overflow Errors I don't exactly know.





The Picture of the Simulation Shows just the Imaginary Part. The Real Part is nearly 0 that means it's about -1 and 1.

Calculation Times@100Mhz Input Clock are about 56µs@128Samples and 23µs@64Samples.
So it's not a pretty fast Implementation and it's not really resource efficient I think. It would be also possible to work with lower Resolution actually I'm using 64bit Fixed Point Numbers (EN25_39).
Resource Usage at 128Samples is:
  • Slice LUT's -> 18988 (29.95%)
  • Slice Registers -> 17050 (13.45%)
  • F7 Muxes -> 4352 (13.73%)
  • F8 Muxes -> 2048 (12.92%)
  • DSPs -> 32 (13.33%)
I think first I will use this implementation to do some work in Software and see some Spectrograms. Also I think I need to implement DMA to my PCIe Design to be more flexible with Software Timings.
But better Implementations for Resource Efficient and faster Fourier Transform would be the Rader Algorithm or the Winograd Algorithm.

In Common.vhd you can find some Constants. These Constants are just the Real and Imaginary Part of a Complex Number with Amplitude 1 and Angle Theta.
Theta is (2*Pi)/unityStep
UnityStep is 0x1 left shifted to the Number of Samples.

Source Code:

Common.vhd

it_fft.vhd

it_fft_tb.vhd

Kommentare

Kommentar veröffentlichen

Beliebte Posts aus diesem Blog

FPGA 50Hz Notch and 1-10Hz Highpass Filter Implementation

Bild
This Weekend I was working on implementing some Filters on the FPGA for my ADS Embedded Project. First I tried it with Fixed Point Arithmetic as I mentioned it in my previous Post. But after reading and trying a bit with it I decided to Switch to Floating Point cause I still have a working Floating Point Adder from a University Project and some thinking about it let's me Implement a working Floating Point Multiplier and the Integer to Floating Point Conversion Units. I don't think that I will release the Units here because I don't know about the licensing. But If you are interested in these Parts send me a mail and I will send you the Code you Need. After I tested them and removing some bugs I started to Implement my Filter. In my University I started to let Matlab create the Filters with the HDL Coder but it doesn't work pretty well and take much resources, because I don't optimized it well in the Settings. But the Simulink Models I created for that we...
Mehr anzeigen

FPGA 50Hz Notch ... Second Version

Bild
Normally I wanted to Show you yesterday my results of the Filter Implementation, but I found no time so I will Show it today. Cause my Electrodes hadn't good contact you won't see a real ECG but you can see how it filters. Orange is the Original unfiltered Signal. Blue is the Signal filtered with the FPGA. Yellow the Signal filtered with Matlab. Matlab and the FPGA uses the same Coefficients so there's much more Noise from the FPGA. I thought I can reduce the Noise with double Floating Point Precision but that won't met my Timings. That let me come back to Fixed Point Arithmetic. So I made this evening a Fixpoint Version I used 25 Bits Precomma and 39 Bits Pastcomma. So in Sum I'm using 64 Bit Vectors and 128 for Multiplication. When I wrote it I used the STD_LOGIC_ARITH and STD_LOGIC_UNSIGNED librarys. When I simulated it first I hadn't get the right results so I thought about it and I realized that Unsigned is the wrong library I had to use ...
Mehr anzeigen

FPGA Notch the third and PCIe Block Diagram

Bild
Today I implemented the Fixed Point Version of my Filter in my IP Core. For the Floating Point Version I were just able to use generate and could implement n Independent Filter for every channel. Cause the Fixed Point Filter is much faster I can Pipeline it through all channels and it's still fast enough. That gives me this total resource consumption of my IP: Slice LUT's -> 2745 (4.33%) Slice Registers -> 5392 (4.25%) F7 Muxes -> 408 (1.29%) RAMB18 -> 0 DSPs -> 16 (6.67%) The Time needed for Filtering 8 Channels is 9.56µs. So with this resource usage I think there is still enough Space to implement a STFT on the Artix 7 ;-) But I won't talk too much about the Filter Implementation. I think it could be interesting to Show you the PCIe Block Diagram and what's important to get it ready. MSI Request -> '0' axi_aresetn -> '1' Utility Buffer: Set to IBUFDSGTE AXI Memory Mapped to PCI Ex...
Mehr anzeigen