Add optional `allow_single_version` Boolean arg to `actx.compile`

[inducer]

Benefits:

• Avoids unexpected repeated compilation
• Avoid dispatch cost to select which implementation to use

Drawbacks:

• Less flexible, of course :)

Naming of the flag is an open question.

cc @majosm @MTCam

[kaushikcfd]

I don't think I understand this. Could you please elaborate? Does this have to do with sub-expressions having the same value but different metadata?
Thanks!

[inducer]

Sorry, I didn't spend enough time explaining. Right now, we have reasonably elaborate dispatch logic between different versions of a compiled function, based on shapes, which arguments are supplied, etc. This flag would assert that only a single version should ever exist, and calling the compiled function a second time with mismatched arguments would be an error. Does that make sense?

[kaushikcfd]

Yep, makes sense. Thanks! Essentially this flag should ensure that length of BaseLazilyCompilingFunctionCaller.program_cache should not exceed 1.

[inducer]

Yup, exactly.

[majosm]

Question: Would the allow_single_version=True case check if the arguments are the same as when it was compiled? Or is it more of an "at your own risk" thing?

[kaushikcfd]

Question: Would the allow_single_version=True case check if the arguments are the same as when it was compiled? Or is it more of an "at your own risk" thing?

It should be an error at this point if len(self.program_cache) > 1.

[majosm]

Question: Would the allow_single_version=True case check if the arguments are the same as when it was compiled? Or is it more of an "at your own risk" thing?

It should be an error at this point if len(self.program_cache) > 1.

Does it still incur the "dispatch cost to select which implementation to use" then?

[kaushikcfd]

Aah fair. I missed that as a constraint. Though I am unsure how to avoid the dispatch cost since we would always need to compute the arg_id_to_descr.

[majosm]

In any case, I can see the value in avoiding multiple compiles. On the user side, though, I'm not sure if it's always obvious enough when it's happening that one would know to intervene. Kind of makes me wonder if the single_version_only=True behavior should be the default? i.e., emit an error if it tries to compile multiple times, unless the caller explicitly says it's OK. Not sure how much stuff this would break, though.

[inducer]

At the very least, it would break the LSRK45 time stepper, which relies on having two different versions at the moment.

https://github.com/inducer/grudge/blob/c9906c9a20bfb5c78733093ccba716d7ac036897/grudge/shortcuts.py#L44-L62

[matthiasdiener]

How about just emitting a warning instead of an exception? This would even work without necessarily changing the API.

[inducer]

IMO, both multi-version and single-version are legitimate uses of the function, unless the user says otherwise.