Tutorial 3: Wideband Spectrometer

Author: Jason Manley and Danny Price

Expected completion time: 2hrs

Introduction

A spectrometer is something that takes a signal in the time domain and converts it to the frequency domain. In digital systems, this is generally achieved by utilising the FFT (Fast Fourier Transform) algorithm. However, with a little bit more effort, the signal to noise performance can be increased greatly by using a Polyphase Filter Bank (PFB) based approach.

When designing a spectrometer for astronomical applications, it's important to consider the science case behind it. For example, pulsar timing searches will need a spectrometer which can dump spectra on short timescales, so the rate of change of the spectra can be observed. In contrast, a deep field HI survey will accumulate multiple spectra to increase the signal to noise ratio. It's also important to note that “bigger isn't always better”; the higher your spectral and time resolution are, the more data your computer (and scientist on the other end) will have to deal with. For now, let's skip the science case and familiarize ourselves with an example spectrometer.

Setup

If you are working on this tutorial at a CASPER Workshop, you should have received separate network connection and log-in instructions. (See Tutorial Setup here). If you are working through this tutorial outside of a workshop, you will need to install and configure the necessary Matlab and Xilinx software tools as well as the most current version of the CASPER libraries.

You should have completed Tutorials 1 and 2 before starting this tutorial in order to get a good understanding of the CASPER toolflow and Simulink designs.

This tutorial comes with a completed model file, a compiled bitstream, ready for execution on ROACH, as well as a Python script to configure the ROACH and make plots.

Spectrometer Basics

When designing a spectrometer there are a few main parameters of note:

• Bandwidth: The width of your frequency spectrum, in Hz. This depends on the sampling rate; for complex sampled data this is equivalent to:

.

In contrast, for real or Nyquist sampled data the rate is half this:

,

as two samples are required to reconstruct a given waveform .

• Frequency resolution: The frequency resolution of a spectrometer, Δf, is given by

,

and is the width of each frequency bin. Correspondingly, Δf is a measure of how precise you can measure a frequency.

• Time resolution: Time resolution is simply the spectral dump rate of your instrument. We generally accumulate multiple spectra to average out noise; the more accumulations we do, the lower the time resolution. For looking at short timescale events, such as pulsar bursts, higher time resolution is necessary; conversely, if we want to look at a weak HI signal, a long accumulation time is required, so time resolution is less important.

Simulink / CASPER Toolflow

Simulink Design Overview

If you're reading this, then you've already managed to find all the tutorial files. Jason has gone to great effort to create an easy to follow simulink model that compiles and works. By now, I presume you can open the model file and have a vague idea of what's happening. The best way to understand fully is to follow the arrows, go through what each block is doing and make sure you know why each step is done. To help you through, there's some “blockumentation” in the appendix, which should (hopefully) answer all questions you may have. A brief rundown before you get down and dirty:

• The all important Xilinx token is placed to allow System Generator to be called to compile the design.

• In the MSSGE block, the hardware type is set to 'ROACH:sx95t” and clock rate is specified as 200MHz.

• The input signal is digitised by the ADC, resulting in four parallel time samples of 8 bits each clock cycle. The ADC runs at 800MHz, which gives a 400MHz nyquist sampled spectrum. The output range is a signed number in the range -1 to +1 (ie 7 bits after the decimal point). This is expressed as fix_8_7.

• The four parallel time samples pass through the pfb_fir_real and fft_wideband_real blocks, which together constitute a polyphase filter bank. We've selected 212=4096 points, so we'll have a 211=2048 channel filter bank.

• You may notice Xilinx delay blocks dotted all over the design. It's common practice to add these into the design as it makes it easier to fit the design into the logic of the FPGA. It consumes more resources, but eases signal timing-induced placement restrictions.

• The real and imaginary (sine and cosine value) components of the FFT are plugged into power blocks, to convert from complex values to real power values by squaring. They are also scaled by a gain factor before being quantised...

• These power values are then requantized by the quant0 block, from 36.34 bits to 6.5 bits, in preparation for accumulation. This is done to limit bit growth.

• The requantized signals then enter the vector accumulators, vacc0 and vacc1, which are simple_bram_vacc 32 bit vector accumulators. Accumulation length is controlled by the acc_cntrl block.

• The accumulated signal is then fed into software registers, odd and even.

Without further ado, open up the model file and start clicking on things, referring the blockumentation as you go.

adc

The first step to creating a frequency spectrum is to digitize the signal. This is done with an ADC – an Analogue to Digital Converter. In Simulink, the ADC daughter board is represented by a yellow block. Work through the “iADC tutorial” if you're not familiar with the iADC card.

The ADC block converts analog inputs to digital outputs. Every clock cycle, the inputs are sampled and digitized to 8 bit binary point numbers in the range of -1 to 1 and are then output by the ADC. This is achieved through the use of two's-compliment representation with the binary point placed after the seven least significant bits. This means we can represent numbers from -128/128 through to 127/128 including the number 0. Simulink represents such numbers with a fix_8_7 moniker.

ADCs often internally bias themselves to halfway between 0 and -1 (ie the center of the range of representable values). This means that you'd typically see the output of an ADC toggling between zero and -1 when there's no input. It also means that unless otherwise calibrated, an ADC will have a negative DC offset.

The ADC has to be clocked to four times that of the FPGA clock. In this design the ADC is clocked to 800MHz, so the ROACH will be clocked to 200MHz . This gives us a bandwidth of 400MHz, as Nyquist sampling requires two samples (or more) each second.

INPUTS

sim_in	Input for simulated data. It's useful to connect up a simulink source, such as “band-limited white noise” or a sine wave.
sim_sync	Simulated sync pulse input. In this design, we've connected up a constant with value '1'.
sim_data_valid	Can be set to either 0 (not valid) or 1 (valid).

OUTPUTS

The ADC outputs two main signals: i and q, which correspond to the coaxial inputs of the ADC board. In this tutorial, we'll only be using input i. As the ADC runs at 4x the FPGA rate, there are four parallel time sampled outputs: i0, i1, i2 and i3. As mentioned before, these outputs are 8.7 bit.

pfb_fir_real

There are two main blocks required for a polyphase filter bank. The first is the pfb_fir_real block, which divides the signal into parallel 'taps' then applies finite impulse response filters (FIR). The output of this block is still a time-domain signal. When combined with the FFT_wideband_real block, this constitutes a polyphase filterbank.

INPUTS/OUTPUTS

As the ADC has four parallel time sampled outputs: i0, i1, i2 and i3, we need four parallel inputs for this PFB implementation.

PARAMETERS

fft_wideband_real

The FFT_wideband_real block is the most important part of the design to understand. The cool green of the FFT block hides the complex and confusing FFT butterfly biplex algorithms that are under the hood. You do need to have a working knowledge of it though, so I recommend reading Chapter 8 and Chapter 12 of Smith's free online DSP guide (at http://www.dspguide.com/). Parts of the documentation below are taken from the block documentation by Aaron Parsons and Andrew Martens.

INPUTS/OUTPUTS

PARAMETERS

power

The power block computes the power of a complex number. The power block typically has a latency of 5 and will compute the power of its input by taking the sum of the squares of its real and imaginary components. The power block is written by Aaron Parsons and online documentation is by Ben Blackman. In our design, there are two power blocks, which compute the power of the odd and even outputs of the FFT. The output of the block is 36.34 bits; the next stage of the design re-quantizes this down to a lower bitrate.

INPUTS/OUTPUTS

PARAMETERS

quant

The quant0 was written by Jason Manley for this tutorial and is not part of the CASPER blockset. The block re-quantizes from 36.34 bits to 6.5 unsigned bits, in preparation for accumulation by the 32 bit bram_vacc block. This block also adds gain control, via a software register. The spectrometer.py script sets this gain control. You would not need to re-quantize if you used a larger vacc block, such as the 64bit one, but it's illustrative to see a simple example of re-quantization, so it's in the design anyway. Note that the sync_out port is connected to a block, acc_cntrl, which provides accumulation control.

INPUTS/OUTPUTS

PARAMETERS

None.

simple_bramm_vac

The simple_bram_vacc block is used in this design for vector accumulation. Vector growth is approximately 28 bits each second, so if you wanted a really long accumulation (say a few hours), you'd have to use a block such as the qdr_vacc or dram_vacc. As the name suggests, the simple_bram_vacc is simpler so it is fine for this demo spectrometer. The FFT block outputs 1024 cosine values (odd) and 1024 sine values, making 2048 values in total. We have two of these bram vacc's in the design, one for the odd and one for the even frequency channels. The vector length is thus set to 1024 on both.

PARAMETERS

INPUTS/OUTPUTS

Even and Odd BRAMs

The final blocks, odd and even, are shared BRAMs, which we will read out the values of using the spectrometer.py script.

PARAMETERS

INPUTS/OUTPUTS

Software Registers

There are a few control registers, led blinkers, and snap block dotted around the design too:

• cnt_rst: Counter reset control. Pulse this high to reset all counters back to zero.

• acc_len: Sets the accumulation length. Have a look in spectrometer.py for usage.

• sync_cnt: Sync pulse counter. Counts the number of sync pulses issued. Can be used to figure out board uptime and confirm that your design is being clocked correctly.

• acc_cnt: Accumulation counter. Keeps track of how many accumulations have been done.

• led0_sync: Back on topic: the led0_sync light flashes each time a sync pulse is generated. It lets you know your ROACH is alive.

• led1_new_acc: This lights up led1 each time a new accumulation is triggered.

• led2_acc_clip: This lights up led2 whenever clipping is detected.

There are also some snap blocks, which capture data from the FPGA fabric and makes it accessible to the Power PC. This tutorial doesn't go into these blocks (in its current revision, at least), but if you have the inclination, have a look at their documentation. In this design, the snap blocks are placed such that they can give useful debugging information. You can probe these through KATCP, as done in Tutorial 1, if interested. If you've made it to here, congratulations, go and get yourself a cup of tea and a biscuit, then come back for part two, which explains the second part of the tutorial – actually getting the spectrometer running, and having a look at some spectra.

Configuration and Control

Hardware Configuration

The tutorial comes with a pre-compiled bof file, which is generated from the model you just went through ( r_spec_2048_r105_2010_Jul_26_1205.bof) Copy this over to you ROACH boffiles directory, chmod it to a+x as in the other tutorials, then load up your ROACH. You don't need to telnet in to the ROACH; all communication and configuration will be done by the python control script called spectrometer.py.

Next, you need to set up your ROACH. Switch it on, making sure that:

• You have your ADC in ZDOK0, which is the one nearest to the power supply.

• You have your clock source connected to clk_i on the ADC, which is the second on the right. It should be generating an 800MHz sine wave with 0dBm power.

The spectrometer.py script

Once you've got that done, it's time to run the script. First, check that you've connected the ADC to ZDOK0, and that the clock source is connected to clk_i of the ADC. Now, if you're in linux, browse to where the spectrometer.py file is in a terminal and at the prompt type

 
 .\spectrometer.py <roach IP or hostname> -b <boffile name>

replacing <roach IP or hostname> with the IP address of your ROACH and <boffile name> with your boffile. You should see a spectrum like this:

In the plot, there should be a fixed DC offset spike; and if you're putting in a tone, you should also see a spike at the correct input frequency. If you'd like to take a closer look, click the icon that is below your plot and third from the right, then select a section you'd like to zoom in to.

Now you've seen the python script running, let's go under the hood and have a look at how the FPGA is programmed and how data is interrogated. To stop the python script running, go back to the terminal and press ctrl + c a few times.

iPython walkthrough

The spectrometer.py script has quite a few lines of code, which you might find daunting at first. Fear not though, it's all pretty easy. To whet your whistle, let's start off by operating the spectrometer through iPython. Open up a terminal and type:

ipython --pylab

and press enter. You'll be transported into the magical world of iPython, where we can do our scripting line by line, similar to MATLAB. Our first command will be to import the python packages we're going to use:

import corr,time,numpy,struct,sys,logging,pylab

Next, we set a few variables:

katcp_port = 7147
roach = 'enter IP address or hostname here'
timeout = 10

Which we can then use in FpgaClient() such that we can connect to the ROACH and issue commands to the FPGA:

fpga = corr.katcp_wrapper.FpgaClient(roach,katcp_port, timeout)

We now have an fpga object to play around with. To check if you managed to connect to your ROACH, type:

 
fpga.is_connected()

Let's set the bitstream running using the progdev() command:

fpga.progdev('r_spec_2048_r103_2010_Jul_26_1205.bof')

Now we need to configure the accumulation length and gain by writing values to their registers. For two seconds and maximum gain: accumulation length, 2*(2^28)/2048, or just under 2 seconds:

fpga.write_int('acc_len',2*(2**28)/2048)
fpga.write_int('quant0_gain',0xffffffff)

Finally, we reset the counters:

fpga.write_int('cnt_rst',1)
fpga.write_int('cnt_rst',0)

To read out the integration number, we use fpga.read_uint():

fpga.read_uint('acc_cnt')

Do this a few times, waiting a few seconds in between. You should be able to see this slowly rising. Now we're ready to plot a spectrum. We want to grab the even and odd registers of our PFB:

a_0=struct.unpack('>1024l',fpga.read('even',1024*4,0))
a_1=struct.unpack('>1024l',fpga.read('odd',1024*4,0))

These need to be interleaved, so we can plot the spectrum. We can use a for loop to do this:

interleave_a=[]
 
for i in range(1024):
	interleave_a.append(a_0[i])
	interleave_a.append(a_1[i])

This gives us a 2048 channel spectrum. Finally, we can plot the spectrum using pyLab:

pylab.figure(num=1,figsize=(10,10))
pylab.plot(interleave_a)
pylab.title('Integration number %i.'%prev_integration)
pylab.ylabel('Power (arbitrary units)')
pylab.grid()
pylab.xlabel('Channel')
pylab.xlim(0,2048)
pylab.show()

Voila! You have successfully controlled the ROACH spectrometer using python, and plotted a spectrum. Bravo! You should now have enough of an idea of what's going on to tackle the python script. Type exit() to quit ipython.

spectrometer.py notes

Now you're ready to have a closer look at the spectrometer.py script. Open it with your favorite editor. Again, line by line is the only way to fully understand it, but to give you a head start, here's a few notes:

Connecting to the ROACH

To make a connection to the ROACH, we need to know what port to connect to, and the IP address or hostname of our ROACH. The connection is made on line 98:

88. fpga = corr.katcp_wrapper.FpgaClient(...)

The katcp_port variable is set on line 16, and the roach variable is passed to the script at the terminal (remember that you typed python spectrometer.py roachname). We can check if the connection worked by using fpga.is_connected(), which returns true or false:

101. if fpga.is_connected():

The next step is to get the right bitstream programmed onto the FPGA fabric. The bitstream is set on line 15:

18. bitstream = 'r_spec_2048_r103_2009_Nov_23_1234.bof'

Then the progdev command is issued on line 101:

110. fpga.progdev(bitstream)

Passing variables to the script

Starting from line 66, you'll see the following code:

from optparse import OptionParser
 
p = OptionParser()
p.set_usage('spectrometer.py <ROACH_HOSTNAME_or_IP> [options]')
p.set_description(__doc__)
 
p.add_option('-l','—acc_len',dest='acc_len',
type='int', default=2*(2**28)/2048,
help='Set the number of vectors to accumulate between dumps. default is 2*(2^28)/2048, or just under 2 seconds.')
 
p.add_option('-g', '--gain', dest='gain',
type='int',default=0xffffffff,
help='Set the digital gain (6bit quantisation scalar). Default is 0xffffffff (max), good for wideband noise. Set lower for CW tones.')
 
p.add_option('-s', '--skip', dest='skip', action='store_true',
help='Skip reprogramming the FPGA and configuring EQ.')
 
opts, args = p.parse_args(sys.argv[1:])
 
 if args==[]:
     print 'Please specify a ROACH board. Run with the -h flag to see 	all options.\nExiting.'
     exit()
 else:
     roach = args[0]

What this code does is set up some defaults parameters which we can pass to the script from the command line. If the flags aren't present, it will default to the values set here.

Conclusion

If you have followed this tutorial faithfully, you should now know:

• What a spectrometer is and what the important parameters for astronomy are.

• Which CASPER blocks you might want to use to make a spectrometer, and how to connect them up in Simulink.

• How to connect to and control a ROACH spectrometer using python scripting.