[discussion aborted due to lack of knowledgeable persons on the call to make any progress]
L1 down-shift by eight
The basic algorithm is that there's a mixer stage, a down-shift by eight, and a 256 tap decimating FIR filter. The input bandwidth is 1500 MHz. The output bandwidth is 187 MHz. One question that springs to mind immediately is why the downshift by 8? The input is a 9.7 fixed point number, and the output is a 12.10 number.
Need UCB to comment on what seems to be a numeric instability
Randy may simulate to verify what is happening
Filter coefficients (MATLAB vs. a normalized design
The coefficients were calculated with the MATLAB filter design tool. I am pretty sure I have a screenshot of the settings I used on my computer at work. I never intended that to be the final filter, just something that would work to test the system. The coeffs are entered into the field of the FIR green block, and then are stored as constants hooked to the appropriate multipliers within the inner workings of the block. If the filter were normalized in some way that the integral across the band for an all pass were 1, then this 1/8th band filter would cut the signal down. But I think the actual gain is less than that. I am sure that the down-shift was not trying to account for that change in power, more likely it was to ensure no overflow in the adders.
Half-band filter in the middle of the L8 bandwidth
The L8 design uses a combination of CIC & half-band filter. The CIC filter has fixed coefficients (all 1's) & bit growth. There are some bit selection (truncation) between stages of the CIC filters to avoid it getting too big. I'm more worried about the half-band filter in the middle now. The coefficients of the half-band filter are calculated using the script (https://github.com/casper-astro/vegas_devel/blob/master/lib/gbt_library/halfband_coeff_gen.m) and stored as constants (values as below).
Need someone from UCB to comment
ADC values seem to be missing in certain circumstances
Why do some of the input ADC values seem to be missing? E.g. look at Mode 4, Scan 14 on: http://www.gb.nrao.edu/~rprestag/vegasPlots/ Although the L1 filtersnap_nogain values are bogus, all of the imaginary values seem to be identically zero. Is this expected? In the l8 modes, the sb1snap values dont look particularly Gaussian. Again, is this a problem?