sr-cdf
commented
7 months ago TL;DR Software rotation is probs fine, but can we have independent RX and TX LOs anyway?
Updating the tone frequency to follow the moving resonator involves reading the RX accumulator, calculating the phase shift since the last reading, looking up how much frequency shift this corresponds to and applying the new frequency to the LO, with an update of the channel selector if necessary.
It is certainly important to correct the phase delay imparted by the system, two reasons come to mind:
-
If the phase wraps between samples it will be difficult to estimate the frequency shift.
-
If the system phase is changing with frequency faster than the resonator phase changes with frequency then it will be harder to track the resonator phase.
As you say, if there is a firmware complex multiplier to apply a rotation on the RX or TX LO (or the accumulator input) to correct for the system phase delay, the value of the rotation is frequency dependent and will need updating updating whenever we set a new frequency. A one time calibration sweep would suffice to inform us of the frequency dependent phase shift values.
Alternatively, we software complex multiply every sample out of the accumulator by the required system phase shift value in the main loop. The value could be pre-computed and stored an array somewhere in memory.
So, which takes up fewer clock cycles? Looking up 1000 formatted phase shift values and writing them to 1000 registers, or looking up 1000 phase shift integer values, and complex multiplying 1000 iq pairs in the software loop?
If software multiply if faster than firmware register reload, then I would agree there is no need for a firmware correction.
Alternatively, (and you may have hinted at this), we could have separate independent LOs for RX and TX, each with its own programmable offset. We can perform software rotation in the main loop, or apply independent offsets to each LO if it turns out to be useful.
The dual LO approach would also allow us sample a wider band around the resonators, by generating a 1GHz tone, receiving at 1GHz and accumulating down to 1Khz, then receiving at 1G+1k, 1G+2k 1G+3k... until we've measured the tone out to across the full resonator bandwidth.
jack-h
The current (v6.3) system uses the same LO to generate tones and mix the received tones. No compensation is made for the travel time out the TX system, through the MKID, and through the RX system.
Compensating this delay as a true-time delay on the LO would be very expensive in RAM. Compensating this delay as a phase rotation on the LO would be less expensive. The group delay first needs obtaining through some calibration sweep, and then an LO-freq-dependent, but constant in time, rotation can be applied to the mixer in either the TX or RX paths. This rotation would need updating whenever the tone frequencies are tweaked, so would marginally slow the update rate.
However, applying this rotation is the same as rotating the data out of the RX accumulators, which can be performed at (at least) the same cadence as tone updates. Therefore (@sr-cdf) what is the tangible benefit of correcting in the firmware (other than it making the plots look prettier, though maybe this is reason enough)?