Discussion: Should functions in lowpp layer be `def` rather than `cpdef`?

[mdboom]

Functions in the lowpp layer are currently declared with cpdef, so they are exposed to both Cython and Python. However, we have always claimed that the Cython interface is not stable, and shouldn't be relied on.

A couple discoveries today are making me wonder whether we want to keep exposing the lowpp layer to Cython at all:

• Calling a cpdef function from Python has an additional 50ns of overhead vs. a def function (the Python layer delegates to an underlying layer and there is additional work done at that boundary).
• cpdef functions can not have their return type annotated, which is a blocker for proper .pyi support for the lowpp layer.

Is Cython access at the lowpp layer providing any value here? Could we just s/cpdef/def/g and get more performance and a more Pythonic experience?

[leofang]

I am not aware of any perf cost! This is interesting...

For driver/runtime/NVRTC: They should be already on def in the old days of cython-gen.

For the rest: This might need a wider cybind user discussion, but I think the lowpp layer was designed with Python-only exposure in mind.