Unexpected behavior for long sequences with APS2

[ranzani]

When running long sequences with the APS2 if the waveform data cross a cache line the prefetcher can't keep up. The experiment proceeds, but the data are random.

Verified this for example with the following test sequence:

[[X(q,length = l), MEAS(q)] for l in np.arange(2e-6,50e-6,2e-6)]

[matthewware]

Just as a note to others, this is an issue with the APS2 waveform cache. For longer pulses that aren't constructed with time/amplitude pairs (constant functions), these pulses can chew through all the on-system waveform memory. You can get around this issue by using waveforms made like in the echo-CR case:

QGL/QGL/PulsePrimitives.py

Lines 358 to 374 in e50941b

	# helper used by echoCR
	def flat_top_gaussian(chan,
	riseFall,
	length,
	amp,
	phase=0,
	label="flat_top_gaussian"):
	"""
	A constant pulse with rising and falling gaussian shape
	"""
	if riseFall == 0:
	p = Utheta(chan, length=length, amp=amp, phase=phase, shape_fun=PulseShapes.constant, label=label+"_top")
	else:
	p = Utheta(chan, length=riseFall, amp=amp, phase=phase, shape_fun=PulseShapes.gaussOn, label=label+"_rise") + \
	Utheta(chan, length=length, amp=amp, phase=phase, shape_fun=PulseShapes.constant, label=label+"_top") + \
	Utheta(chan, length=riseFall, amp=amp, phase=phase, shape_fun=PulseShapes.gaussOff, label=label+"_fall")
	return p._replace(label=label)

Long term, we're hoping to have some checks at compile time to make sure you aren't overflowing the waveform memory.