Compute translation_classes_data if not given

[isuruf]

Adds a _disable_translation_classes hidden option so that we can check the non translation classes kernel.

[inducer]

Thx!

[alexfikl]

Why does this work? Doesn't the queue get destroyed when it exits the context manager leaving SumpyTranslationClassesData holding an invalid reference?

[isuruf]

Yes, the queue completes and gets destroyed, but the data itself doesn't get destroyed. i.e. the data has queue=None. When a new queue needs this data, they are accessed with .with_queue(queue)

[alexfikl]

sumpy/sumpy/fmm.py

Lines 244 to 249 in a2c830a

	def __init__(self, queue, trav, is_translation_per_level=True):
	# FIXME: Queues should not be part of data.
	self.queue = queue
	self.trav = trav
	self.tree = trav.tree
	self.is_translation_per_level = is_translation_per_level

Hm, but the constructor doesn't do any work. The actual work is being done when calling some of those memoized methods below, which should be after the context manager exits.

[alexfikl]

The tests don't seem to crash and burn, so this is probably just some misunderstanding on my side. Hm..

[isuruf]

Hmm, you are right. I guess since the constructor keeps a reference to the queue, it doesn't get destroyed.

[inducer]

I'll try to come up with a way to make PyOpenCL warn if a reference to the queue continues to be held past __exit__.

[alexfikl]

That would be really cool!

[isuruf]

I changed the implementation to build the data objects eagerly instead of lazily.

[alexfikl]

Can't speak for the eager vs lazy, but the queue stuff makes sense to me now 👍

[inducer]

inducer/pyopencl#561

[inducer]

LGTM, thanks!