mirror of https://github.com/m-labs/artiq.git
69 lines
1.8 KiB
ReStructuredText
69 lines
1.8 KiB
ReStructuredText
FAQ
|
|
###
|
|
|
|
How do I ...
|
|
============
|
|
|
|
override the sysclk frequency of just one dds?
|
|
----------------------------------------------
|
|
|
|
Override the parameter using an argument in the ddb.
|
|
|
|
organize parameters in folders?
|
|
-------------------------------
|
|
|
|
Use gui auto-completion and filtering.
|
|
Names need to be unique.
|
|
|
|
enforce functional dependencies between parameters?
|
|
---------------------------------------------------
|
|
|
|
Use wrapper experiments, overriding parameters of arguments.
|
|
|
|
get rid of `DbKeys`?
|
|
--------------------
|
|
|
|
`DbKeys` enforces valid parameter/argument names, references
|
|
keys in pdb and hints at metadata on how values can be retrieved.
|
|
|
|
write a generator feeding a kernel feeding an analyze function?
|
|
---------------------------------------------------------------
|
|
|
|
Like this::
|
|
|
|
def run(self):
|
|
self.parse(self.pipe(iter(range(10))))
|
|
|
|
def pipe(self, gen):
|
|
for i in gen:
|
|
r = self.do(i)
|
|
yield r
|
|
|
|
def parse(self, gen):
|
|
for i in gen:
|
|
pass
|
|
|
|
@kernel
|
|
def do(self, i):
|
|
return i
|
|
|
|
create and use variable lengths arrays?
|
|
------------------------------------------------
|
|
|
|
Don't. Preallocate everything. Or chunk it and e.g. read 100 events per
|
|
function call, push them upstream and retry until the gate time closes.
|
|
|
|
execute multiple slow controller RPCs in parallel without loosing time?
|
|
-----------------------------------------------------------------------
|
|
|
|
Use `threading.Thread`: portable, fast, simple for one-shot calls
|
|
|
|
write part of my experiment as a coroutine/Task/generator?
|
|
----------------------------------------------------------
|
|
|
|
You want to write experiment preparation (`__init__()` or `build()`)
|
|
or analysis (`analyze()`)
|
|
|
|
No. That would make reusing your own code in sub-experiments difficult and
|
|
fragile.
|