docs: Add docs for RTIO SED sequencing (#1461)

Signed-off-by: Charles Baynham <charles.baynham@npl.co.uk>
This commit is contained in:
charlesbaynham 2020-06-15 17:43:34 +01:00 committed by GitHub
parent 1a17d0c869
commit ce7e92a75e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 2 additions and 1 deletions

View File

@ -123,7 +123,7 @@ Sequence errors
---------------
A sequence error happens when the sequence of coarse timestamps cannot be supported by the gateware. For example, there may have been too many timeline rewinds.
Internally, the gateware stores output events in an array of FIFO buffers (the "lanes") and the timestamps in each lane much be strictly increasing. The gateware selects a different lane when an event with a decreasing or equal timestamp is submitted. A sequence error occurs when no appropriate lane can be found.
Internally, the gateware stores output events in an array of FIFO buffers (the "lanes") and the timestamps in each lane must be strictly increasing. If an event with a decreasing or equal timestamp is submitted, the gateware selects the next lane, wrapping around if the final lane is reached. If this lane also contains an event with a timestamp beyond the one being submitted then a sequence error occurs. See `this issue <https://github.com/m-labs/artiq/issues/1081>`_ for a real-life example of how this works.
Notes:
@ -131,6 +131,7 @@ Notes:
* Configuring the gateware with more lanes for the RTIO core reduces the frequency of sequence errors.
* The number of lanes is a hard limit on the number of simultaneous RTIO output events.
* Whether a particular sequence of timestamps causes a sequence error or not is fully deterministic (starting from a known RTIO state, e.g. after a reset). Adding a constant offset to the whole sequence does not affect the result.
* Zero-duration methods (such as :meth:`artiq.coredevice.ttl.TTLOut.on()`) do not advance the timeline and so will consume additional lanes if they are scheduled simultaneously. Adding a tiny delay will prevent this (e.g. ``delay_mu(self.core.ref_multiplier)``, at least one coarse rtio cycle).
The offending event is discarded and the RTIO core keeps operating.