Temporal benchmarks

[bruno-f-cruz]

Consider the best clock is the BioData device, ideally from the hardware timestamped events (e.g. SynchPulse or 'Set' stream).

Specifically:

• [x] Microphone:
  • Benchmark with the pulse that was sent to the mic ADC which relayed the synchpulse. Since each buffer is timestamped in CPU we can compare these two events and have a sense of the delay/jitter between the syncpulse (given in HARP HW) and when it arrives in Bonsai (given by the latest harp timestamp)
• [x] TinkerForge
  • Benchmark using the AnalogInput bricklets since it is receiving the synch pulse from Harp. Find when the pulse occurs and compare the hw timestamp with the latest harp timestamp in software. NB: There is an assumption that the delay across bricklets is comparable.
• [x] UBX
  • This one is a bit trickier since the ubx hardware "buffers" the event and sends it at 1HZ. As a result I think the best metric is to simply compare the clock drift using the syncpulse we are sending to the UBX and logging. OR comparing the delay between each PPS message and the corresponding toggle in the output line that we are reading in one of the digital inputs of the BioData board
• [x] Pupil labs
  • We will start by looking at the burned IR LED in the camera and go from there. Unfortunately, we don't have a good way to get that hardcoded information translated into an LSL event. We are taking for granted that all the files logged by the software are aligned with the logs provided by LSL, which might not be the case. In the future, it might be better to pack all the data (e.g. Image, X, Y, etc..) in a single packet that guarantees synchrony. After that, the current benchmark strategy should work.
• [x] Adafruit Accelerometer
  • Assuming the events are well aligned (trigger and serial response) should be as simple as subtracting the two columns (hw_time - software time). That said, there is not any reason why we should not always use the latter...
• [x] Empatica
  • We will use the photodiode that emits light when the "Tag" event is also sent via the server. We will then compare the timestamp of the event given in CPU with the photodiode signal given from the Biodata device

[bruno-f-cruz]

TinkerForge benchmark (f8e5b53):

[bruno-f-cruz]

Empatica benchmarks (in 81e57cf) Processing algorithm (some negative latencies were removed because the algorithm fails to detect very slow LED ramps, ~ 1% fo all trials):

[bruno-f-cruz]

Acceleromter benchmark (in 34a0fcd2d64054fdfba39b0234b5e8d175401f17):

[bruno-f-cruz]

For the UBX, we can check how does the difference grow in the respective clocks of the digital input event injected in the UBX board. We have made the decision of aligning on the first pulse and let the error propagate.

[bruno-f-cruz]

The best we can do for the pupil right now is to assume a stable frame rate (200hz) and calculate the drift by resetting on the first burned LED event. This results in roughly 1400frames (or 7 seconds) missing after ~20min of recording.

[bruno-f-cruz]

Microphone: The difference in the timestamp given in software to a buffer where we detect the TTL and the corresponding hardware-timestamped ttl: