Skip to content

Synchronized start

DtolManager.start_synchronized(names) starts several subsystems together via the DT-Open Layers simultaneous-start primitives, so their first samples begin within one sample period of each other — tighter than a software start loop can achieve.

What it does

It runs the four-step SDK sequence (and always releases the list afterward):

olDaGetSSList            # obtain a simultaneous-start list for the board
olDaPutDassToSSList × N  # add each task's subsystem
olDaSimultaneousPreStart # arm every subsystem
olDaSimultaneousStart    # start them all on one trigger
olDaReleaseSSList        # always, even if a step raised

Usage

import anyio

from dtollib import AnalogInputVoltage, CounterEdgeCount, DtolManager, TaskSpec
from dtollib.backend.dataacq import DataAcqBackend


ai_spec = TaskSpec(name="ai", board="DT9806(00)",
                   channels=[AnalogInputVoltage(physical_channel=0)])
ct_spec = TaskSpec(name="ct", board="DT9806(00)",
                   channels=[CounterEdgeCount(physical_channel=0)])


async def main() -> None:
    backend = DataAcqBackend()  # one shared backend → one HDRVR namespace
    async with DtolManager() as mgr:
        await mgr.add("ai", ai_spec, backend=backend)
        await mgr.add("ct", ct_spec, backend=backend)
        await mgr.start_synchronized(["ai", "ct"])
        # both subsystems are now RUNNING, started together.


anyio.run(main)

Scope and constraints

  • Single board. Every named task must target the same board. Cross-board coordination over the Sync Bus (olDaSetSyncMode) is not yet implemented.
  • One shared backend. Pass the same backend= to each add(). The SS-list is built from one HDRVR's subsystem handles; tasks on separate backend instances live in separate handle namespaces and cannot be coordinated. The manager raises DtolValidationError if the named tasks don't share a backend.
  • Tasks must be committed, not started. DtolManager.add() commits during registration, so single-value and counter tasks are ready immediately.
  • Continuous-AI tasks must complete their record() register → queue → commit sequence first. Open the recorder with autostart=False, let it wire the §12.3.2 callback bridge and commit, then call start_synchronized to begin acquisition in lockstep with the other subsystems. A first-class record_synchronized() helper is deferred (OQ-5c).

Verifying alignment

On the bench, capture both subsystems through a RawCountsSink and compare the first-sample timestamps — they should fall within one sample period. The fake backend models the state transitions (PRESTARTED → RUNNING) and the ordering invariants (put before pre-start, pre-start before start) so the sequence is unit-tested without hardware.