pdev hardware (jeff mock's spectrometer)

Todo:

• 17jun08:
  
• see if the 1342,1358 is an image of the aerostat.
  
• the birdies across the band when 1290 or 1350 are strong. See if pfb overflow.
  

History:

• 23jan23: b3s1t0  pdev-118. failed.
• had fpga not configured, reset box, fpga configured,
• but pdevinfo.sc  could not read plinth version.
• when adjusting mock power , got prun dma error.
  
• 28dec22: b5s1g0 pdev-103 hands. fpga not configured. resetting box, you can configure it , then a minute later it says not configured.
• replaced with pdev-126.
  
• 08feb22: b5sxG0 : pdev112 .. no dma connection to fpga chip2 (high freq band).
  
• pdevinfo.sc gave 0's or id,versions, etc.. fpga1(low freq) reads id, versions ok. .
• . replaced with pdev 103 (what was the problem with pdev103).
  
• 19feb20: pdev103  replaced..
  
• 08jun08:
•  Hires mode hr_shift was set to scale as 10 - log2(sqrt(decimation)) which is correct for noise. Testing (26sep07) showed that the scaling was ok, but the level was 3.75 too low. After looking at the  fpga code i see a factor of 2 scaling. I updated the pnetbld default scaling to be 12 - log2(sqrt(decimation)). The rms should now  come out close to the input a/d values. Need to check that this works.
  

Spectrometer specifications:

A/D and mixer Specs:   (top)

PDEV Data formats: (top)

• Header:
• 1024 byte header at start of first file of integration. If scan spans multiple files then all files after the first do not have the 1024 byte header.
• 32 int*4 word general header
• sp1 header of int*2  * N words describing the sp1 parameters.
  
• The header is normally output prior to datataking start.
• spectral Data.
  
• a=polA*2, b=polB*2, u=stokesU (2*RE(BA*), v=stokesV (2*Im(BA*))
  
• Normal format is : a1,b1,u1,v1 .. a2,b2,u2,v2.... for nchan  channels.
• 32bit stokes: i1,q1 i2,q2.. to nchan.. then u1,v1 u2,v2 .. to nchan
• After the 1,2,or 4 spectra of a dump, there is an 8 byte header that holds:
• seqNumber (16 bits)
• fftaccums (16 bits) done for this dump
• calon (4 bits)
• adc,pfb,vshift,accum,ashift overflows .. all 4 bit values.
  
• out of the spectrometer the data is in bit reversed order. The program bpInp will bit reverse it.
  
• .pdev file will have the above order .. a1,b1,u1,v1 etc...
• Time domain data: let I,Q be the inphase and quadrature samples, A,B be polA and polB  then:
  
• I0A,Q0A,I0B,Q0B .. I1A,Q1A,I1B,Q1B  
  
• If only 1 pol then order is the same, just remove the unused pol.
• There are no 8 byte record headers in each block of data. Everything is data.
  
• Programming notes:
• the pnet program sends the 32 word header plus the sp1 header (2*56 register values) to psrv via prun
• psrv receives the header msg node and passes it on to bpInp. but rather than saying the data length in the msgnode is (32*4+2*56=240),it tells bpInp that the node datalen is 1024 bytes (since the allocated buffer is much bigger than that). So the bytes after the 240 of the header are not cleared before pushing them through the pipeline (it uses whatever was in the buffer). bpInp pushes the header node to bpOut. bpOut outputs the header using the number of bytes specified in the node size variable in the node (1024 bytes in this example).
  

Clock frequencies for integral dump times:  (top)

•  - in a column implies that the selection is not allowable (eg 1 second dump of 256 lenfft is larger than 64K which is the size of the dump register.
• the clock freqeuncies fora given fftlen will work for all fft lengths smaller. The table has differing values only because the ones lower in the table are closer to 170 MHz.

Using multiple bandwidths in 1 observation

    It is possible to use different bandwidths  during a single observation. To do so, the integration time and wall time of the different bands must be the same. These times are computed from the parameters that define the setup.

Terms:

• clk :   complex sampler rate.
  
• fftlen - length of fft used
• nfft    - number off ffts summed  for 1 integration time.
• dec    - decimation of sampling. (high res mode). 1 to 1024. The output bandwidth will be the clockFreq/dec
• fftTm - time for 1 fft: fftlen*dec/clk
  
• itm    - integration time. input by user.
• It must be a multiple of the time for 1 fft (decimation*fftlen/clk).
  
• The setup program checks each band sequentially. If a band needs a different integration time, it will modify the requested time.
  
• The integration time used will be the last one computed.
• dcnt  - is 0 for  regular observing. 1 for full stokes.
  
• full stokes requires 1 extra fftTm to dump the data. no integration occurs.
• wtm  - wall time.
• integration time + dcnt*fftTm
  
• full stokes requires 1 extra fftTm to dump the data. no integration occurs.

Equations:

• itm - requested by user. cima will round to an even number of fft times
  
• Nfft1*dec1*fftlen1/clock
  
• nfft = itm/(fftTm) = itm/(dec*fftlen/clk)  rounded to closest integer.
  
• done separately for each band.
  
• wtm =   (nfft + dcnt)*fftTm =[ itm/(dec*fftlen/clk) + dcnt]*dec*fftlen/clk
  

Restrictions:

• bands must have same integration times:
• N1*dec1*fftlen1 = N2*dec2*fftlen2
• It helps for the decimations to divide evenly into one another.
  
• wall time (wtm) equal for both bandwidth configurations.
• let 1,2 be the two configurations. then
• wtm1=(nfft1 + dcnt)*fftTm1 = wtm2=(nfft2 + dcnt)*ffttm2
• substituting for nfft,ffttm gives
• Non stokes data:
  
• dcnt=0,  wtm=itm, so itm1=itm2
  
• nfft1*dec1*fftlen1=nfft2*dec2*fftlen2
  
• dec1/dec2  = (nfft2*fftlen2)/(nfft2*fftlen1) = must be an integral value
• If fftlen1=fftlen2 then
• dec1/dec2= nfft2/nfft1
• So if dec1 is divisible by dec2 , then things will work ok.
• For stokes data:
• dcnt=1, wtm = itm + ffttm
  
• dec1*(nfft1+1)*fftlen1=dec2*(nfft2 + 1)*fftlen2
• After substituting for nfft2
• fftlen1*dec1 = fftlen2*dec2
• So the only way for stokes to have different bandwidths is to shorten the fftlen by the same amount you decrease the bandwidth.

Examples:

• Non stokes:
• dec1= 20, dec2 = 1
  
• dec1=6, dec2 = 12
  
• stokes data:
• fftlen1=4096, dec1=250
• fftlen2=8192, dec2=125
• fftlen1*dec1 = fftlen2*dec2..

Summary:

• Non stokes data
• dec1, dec2 are divisible
• Stokes data:
• dec1*fftlen1 = dec2*fftlen2
  

Compression levels.   (top)

• 17jun08: 16 bit galfacts setup shows 1290 radar compressing.
  
• 26apr08:bbmixer compression levels with a sine wave input.
• Shows that we can cover most of the A/D range (to about 1800 counts) before compression starts).
• 24apr08:A/D rms gain vs mixer gain and compression. On sky and noise source.
• Used noise from sky and if2 noise source
• Measured gain level for optimum sampling on sky.
• Found maximum A/D rms on sky (max gain)
• Using if2 noise source measured compression levels vs gain and a/d rms.
• apr08: Resistors in A/D removed pkToPk full scale  is now .77 volts (was 1.6)
• 14mar08: Signal levels in the bbm injecting a sine wave at the bbm input.
• Measured A/D sigmas vs sinewave power showing compression onset
  
• Computed power levels in mixer chassis at compression onset.
• 14mar08: A/D Signal levels injecting sine wave at A/D input
• Measured A/D rms vs input voltage.
• Shows input level needed to generate full scale sine wave.
• Configuration notes for power levels/dynamic range.

Blanking (top)

• Blanking is done at the individual fft level. On each fft computed, the decision is made whether to include the fft in the accumulation or not.
• A 172 Mhz bw, 1024 fft uses 5.9  usecs of data (excluding the polyphase filter overlap) so blanking can be done every 5.9 usecs.
  
•  When blanking is done, all spectra/crossspectra of an fft are blanked as a unit.
• The spectrometer returns the number of ffts accumulated on every integrated spectra read out.
• Blanking can be done with an external ttl signal or using Saturation of the ADC.
• For external blanking, if any portion of an fft occurs while the blanking signal is true, then the fft is not included
• For adc blanking, If the adc clips for N samples within an fft then the fft is not included. N is programmable. Typically we set it to 1.

Controlling the attenuators (top)

• IF in (polA) or (polB) Call it: IN1
  
• split signal into high IF band, Low IF band IN1_H, IN1_L
  
• attenuate each band separately IN1_H, IN1_L
  
• complex mix each band to base band.
• IN1_H -> IN1_H_bb
  
• IN1_L-> IN1_L_bb
• split each base band signal to go to group 0, group 1 (primary and commensal observers).
• So each Beam has two pols, and two bands --> 4 attenuators to control
• The attenuators are controlled through two serial ports on the back of the group0 spectrometers: b0Sxg0,b1Sxg0,.. b6Sxg0.
  
• The group 1 spectrometers do not have connections (via their serial ports) to control the attenautors.
• Looking from the back of each spectrometer:
• Left serial port: POLB  /dev/ttyS0
• Right serial port:polA  /dev/ttsS1
• Jeff wrote a small program: pmix-ctl. It runs on the spectrometer box and talks to the attenuators over the serial link.
• an example query would be:
• pmix-ctl --get --port=/dev/ttyS0 (polA) ..
  
• This would return:

Version okay = 0x01
Mixer serial = 0001
Lo band attenuator = 21
Hi band attenuator = 23
Lo band level = 0
Hi band level = 5

• The attn values are actually db steps of gain...
• assume there is a 31 db amp before the attenuator
• the attenuator can then change the gain from 0: (all atten in) to 31 db gain (no atten in)

Adjusting the power levels

• The bbm takes an IF signal, sends it through the attenuators,  and then splits it to go to the two separate groups of spectrometers (bNsNg0 and bNsNg1).
  
• There is only one attenuator for an input band that goes to the two  spectrometer groups.
• When adjusting the power levels, the program uses the values read back from group 0 spectrometer.
• If the gain after the attenuators varies between group 0 and group 1, the power level adjustment for group 1 could be off by this gain difference.

• prun.c (in the spectrometer) routine: prun_sigstat
  
• each call to this routine reads 1 million samples and computes the sum and sum of squares. These values are returned oto the caller (pnet). The computation is done as integers.
  
• pnet perl routine: routine pnet_sigsta:
• This calls prun_sigstat 4 times (for a total of 4 million points). It then computes the mean and the rms.
• The rms computation does not subtract off the mean value.
  
• cal.conf file for digitizer offsets.
• the file /usr/local/pdev/etc/cal.conf (on pdevs1) holds the digitizer offset for each a/d. It can be generated using pnet --sigcal=filename.
  
• It holds an offset in digitizer counts for each a/d. The spectrometer (in the fpga) subtracts these values from the a/d reading  before passing it on to the digital lowpass filter and polyphase filter bank.
  
• jeff has included a scaling option to adjust the gain between various digitizers but it is not currently being used.
  

1. establishes communications with the mixer chassis via an rs232 link
2. Reads the current attenuator settings.
3. Reads the current rms and mean values via the pnetsig program
1. It starts pnetsig and talks to it via a pipe. pnetsig runs pnet --sigstat to continually read the rms/mean values.
2. running pnetsig continually gets around the 3 second startup time for pnet.
   
4. Computes the attenuation needed to set the rms to 30 counts
5. Sends the new attenuator values to the mixer chassis.
6. reads the new values of the rms and mean
7. exits

• During the first rms reading then the attenuator values could be set wrong.
• During the second rms reading (after the settig). The attenuator values would be set correctly but the readback will be in error.
• When reading the data, 4 million points are read in 4 separate 1 million point blocks (with a slight delay between each block). 1 million points take 1e6/172e6 = 5.8 milliseconds. 4*this is about 30 milliseconds. Radar rotating at 12 seconds take about 80 milliseconds to rotate thru the observatory.

• svn/pdev/pdev/dataking: pnet.phil, pdevattn, pnetsig.
• pdevs1:~pdev/pdev/sw/src: prun.c, pmix.c \

Rfi issues.  (top)

• Images of birdies are reflected about the center of the band.
  
• The BBMixer is complex so we need an accurate 90 degree phase and amplitude match between the I,Q samples.
• When a radar compresses the system (bbm or a/d) the phase is most likely screwed up and the amplitude is definitely wrong  so the image rejection is not going to work and the birdie will also appear reflected about the center of the band.
• This is seen in the aerosta, 1290 remy, t and the faa radar. It will only happen when the radar compresses in the mixer or a/d.

• 17sep08: Using ADC blanking to suppress radar harmonics.
  
• 11jun08: faa radar saturates a/d causing spectral harmonics.
  
• jan07: pshift values that work with the FAA radar.
  

Parameters:  (top)

• ALL:
• 05dec08: setting the registers for 16 bit sampling.
• 12oct07: A/D to output: the computational pipeline.
• 
  
• HR_SHIFT - high res mode upshisft after filter
• 20may10: test decimations 1-20 showing power,voltage, jumps, and bandpass shape.
  
• 080608: changed default hr_shift from 10-alog2(sqrt(dec)) to 12-alog2(sqrt(dec)). 26sep07 found rms 3.75 too low. In fgpa code (dlpf) found factor of 2. So correct it to 4 so output rms similar to input rms.
  
• 26sep07: location of noise rms bits in DLPF processing (setting hr_shift).
• PSHIFT - shifting in the butterfly stages
• 121003: How pshift is computed for datatking.
• pnetctl allows for pshift to be passed in for each gxN. but cima currently does not pass it in
• Auto computation:
• Done in pnetbld cmpPshift() if pshift not defined in pnetctl (via cima)
• nshiftsTot=alog2(fftlen)
• nshiftsForce=0 (set in pnetbld)
• nshiftsDif=nshiftTot-nshiftForce  .. how many shift available after the forced shifts
• phshift= int (2^nshiftsForce - 1)  .. this puts nshiftsForce 1's in lower part of pshift
• pshift= pshift < nshiftsDif            .. most sigbit is first butterfly. move nforceshift bits up
• lowshift=  (nshiftsDif %2) == 0)? 0x55555:0xaaaaa..alternate shifts. first bit after nforcebits is no shift
• pshift=  (pshift | lowshift  ) & ( 2^nshifts -1) .. or in lower shifts. mask by total number of bits in fftlen.
  
• Possible mods:
• let cima pass in pshift for each gxN
• 0--> use old old auto mode .. use for large radars in lbw
• -100-> use every other butterfly to shift. first butterfly is a shift
• -101 add one extra shift starting on the left
• -102 add two extra shifts starting on the left
• - ... etc
• -99 add one extra shift starting on the right (last butterfly)(
• -98 add twi  extra shifts starting on the right
• etc..
  
• -200 -> use every  other butterfly to shift (first butterfly is no shift
• -201 add one extra shift starting on the left
• -202 add two extra shifts starting on the left
• - ... etc
• -199 add one extra shift starting on the right (last butterfly)(
• -198 add twi  extra shifts starting on the right
• etc..
  
• 
  
• 26aug08: overflow in the PFB butterfly stages from a sine wave.
• 08jul08: spectral bandpass shape and spectral jumps with low a/d rms values
  
• 27jun08: table of pshift vs fftlen for various setups.
  
• 17mar08: spectral bandpass shape vs A/D and pshift  levels
• DSHIFT - down shift before accumulation
• 10oct07: checking how dshift affects the spectra.
• ASHIFT - upshifting accumulator before packing.
• 05jun08: spectral amplitude and rms vs ashift for 16 bit sampling. polA,polb, stokes U,V.
  
• 01mar08: spectral amplitude and rms vs ashift for 8 and 16 bit packing (polA,B only)

bandpass shapes:

• oct13: digital low pass filters
  
• 08jul08: using if2 noise source a/d rms=80, a/d rms=30, bw=172 MHz, 8k,4k,2k,1k,512 lenfft's.
• 02may08: simulating the gain and bandpass shape of the low pass digital filters.

Spectral mean and rms values.

• 05dec08: mean and rms (by channel) of spectra vs fftlen and sampling time when using 16 bit sampling.
  

Testing voltage sampling mode:

• 180517: disc i/o fails with 40mbyte/sec voltage sampling.
• 31oct13: testing  4 bit time domain radar sampling
• 11sep11: measured rms vs adjpwr value for 8bit time domain sampling (.ps) (.pdf)
• frame 1,2 8 bit rms for various adjpower values (327 receiver).. showing bm0,1,2,3
• frame 3,4 8 bit rms for various adjpower values (lbw reciever).. showing  bms' 0,1,2,3
  
• 11sep11: time domain sampling of 1713+0747
• 23mar11: time domain mode. 10 Mhz bandwidth looking at pulsars.
• 31jan11: psr0301+19 at 327 and 1410 Mhz
  

Examples of various datataking configurations:  (top)

GalFacts

• 4096 channels across 172 Mhz
• 1 millisecond sampling
• full stokes, 16  bit output spectra.
• The parameters are normally set to:
• HR_SHIFT - 0 since hires not used
• PSHIFT:  0xffa
• ASHIFT: polA,polB: 14,14. Stokes U,V: 15,15

• 17jun08: Using alfa. az, za sitting at 270, 10. Take data for 12 seconds (1 radar revolution).    

Measurements in chronological order:  (top)

Configuration notes for power levels/dynamic range.

Misc:

Spectrometer fan replacements:

Pwr consumption measurements 09aug07:  (top)

• acvolts 118
• fileserver alone: 2.8amps=330W
• fileserver + 1 box not running: 3.85amps=455 W
• fileserver + 1 box running 170MHz clock, i/orate 170Mb/sec: 4.2Amps=490 W
  

  fileserver alone	330W
  box alone. not running	125W
  fileserver and box running @170MHz with 70Mb/sec I/O	490W
  (running box+i/o) increase over not running	35W

GPIO General purpose i/o bits.   (top)

• B0: 1 pps input. Selected using the pnet.conf dump parameter ppssel 0   (for bit 0)
  
• B1: Cal input bit. If this bit is hi during a dump then the st.calon status bit will be set. The value comes from the calport multiplexor. Bit selection via pnet.conf [gx] parameter calsel.
  
• B2: Blanking bit. When this bit is high, the current fft will not be accumulated. Set via pnet.conf [gx] parameter blanksel
• B3: Cal output. This bit will be sent to the pdev cal selector and then sent to the calport multplexor to drive the calttl signal. The bit is selected via the pnet.conf [dump] keyword gpoe 0x08

Documentation:  (top)

<- page up
home_~phil