FFT

[Weevil]

Hi all,

i started to do a FFT from exact one sine wave with the XFFT(9.0). This works for the 32bit input signal, but for the 16bit input signal it seems not to work. By doing the FFT i would expect one peak in the frequency spectrum like i get in in the following picture:

XFFT settings:

- Architecture Choice -> Pipelined, Streaming I/O

- Transform Lenght -> 256

- Scaling Options -> Scaled

- Input Data Width -> 16 -> input signal from DDS 32bit

- Input Data Width -> 8 -> input signal from DDS 16bit

For generating a sine wave i use the DDS compiler at a Superious Free Dynamic Range of 96 dB and Frequency Resolution 100 Hz. The Phase Increment is in Streaming mode and set with a constant block to 4096. One sine wave is generated by 256 samples now (number_0 x2).

If the settings are switched now for a 16bit sine input signal the FFT works not like i would expect. Maybe anyone has an idea?

Thanky for any response!

[D@n]

@Weevil,

I'm struggling to understand your picture.

Can you explain for me what it is you are plotting?  What is going into the FFT, and what is coming out?

Even better, can you create a more complete trace showing the input and output FFT signaling as well?

I'm hoping that would make your question easier to follow.  Thanks!

Dan

[Weevil]

@D@n

- "ddsAnalog" is the output signal of the DDS core (displayed as signed decimal) and is the input for the FFT core.

- "FFT_RAW" shows the output of the FFT core (signed decimal).

-> the sourroundings of the ddsAnalog and FFT_RAW signals are on both pictures the same, only the bit size is different.

 

- "m_axis_data_last_0" shows the tlast of the FFT core.

- "number_0" is a counter only to control the DDS output lenght

I hope this helpts to easier follow my question.

 

.....actual i am most confused about the data input size. If i set the XFFT core input size to 16bit i need a 32bit signal input, is this right?

[D@n]

@Weevil,

So then ... why would the output of the FFT look like a ramp?  That doesn't make sense.

I guess I'm struggling to understand why your diagram has this ramp in the first place.  What's its purpose?

Dan

[D@n]

@Weevil,

As to the data input size: yes, a 16'bit FFT input will include both imaginary and real components.

Not sure if the LogiCore FFT offers a real FFT implementation.  Were it to do so, the input would only be 16'bits for a 16'bit FFT.  The output would then alternate between a 16-bit real and a 16-bit imaginary value.  I'd have to dig into the spec's more to know if this is even a possibility for the LogiCore FFT--it's certainly a possibility for an FFT in general though.

Dan

[Weevil]

@D@n

Thank you so much for your reply!

The reason for the ramp is a wrong figure in vivado. As you can see in the attached picture the values before the FFT-peak are 0. If i switch to the Architecture Choice -> Automatically Selected this doesn't happen, but the problem is the same.

So if it is the input i can try to connect a 16bit sine wave generated by the DDS to the 32bit input of the XFFT (XFFT Input Data Width = 16).

Would the following verilog code work for such small block or do i have to do something additional?

module bit_wandler(
    input [15:0] val_in,
    input clk_in,
    output reg [31:0] val_out
    );    

always @(posedge clk_in) 
begin
  
    val_out = val_in;
  
end
    
endmodule

 

[xc6lx45]

It was signed number format, wasn't it? I think this

   val_out = val_in;

doesn't work because the 16-bit word needs to be sign-extended to 32 bits.

If I write it out without tricks (not use the "signed" Verilog keyword) and using combinational logic for simplicity it is

wire [31:0]     val_out = {{16{val_in[15]}}, val_in[15:0]};

In other words, the unused high bits are filled with the MSB (sign bit).

Example:

1'b1 is -1 in 1 bit

8'b11111111 is also -1 in 8 bits

32'b11111111111111111111111111111111 is still -1 but in 32 bits

Try adding 1 to either (-1 +1 =0) and see for yourself what happens... theory.

 

 

[Weevil]

@D@n and @xc6lx45

I tryed it and it looks like its working.

and...

Does the result looks right for you?

@xc6lx45 transfer 16bit -> 32bit, both possibilities are providing the same result.

 

[xc6lx45]

Looks plausible, given ...

positive frequencies : exp(i x); negative frequencies: -exp(- i x)

I haven't read the FFT IP documentation but I suspect you need to split each e.g. 32-bit output bin into 16-bit real and imaginary part

[Weevil]

@xc6lx45

Thank you very much!

Yes i know, this i will do in another function after transfering the data to my computer.