Fully (almost) type tools.py

[alexfikl]

Inspired by all the new errors in #616, this adds types to pretty much everything in tools.py (on a bit of a best effort basis, since some of them may be too strict). Some small-ish parts are still missing:

• type checking for get_gl_sharing_context_properties, mostly due to missing the OpenGL bindings.
• np.dtype doesn't seem to like taking a dict (in match_dtype_to_c_struct), not sure why.
• mako still isn't typed and causing some small errors.
• Some issues around cl.array.Array.data not having the right type. This might just need some checking to make sure it's ok.

[inducer]

Found a way that avoids all the casting.

Btw, given the choice between cast and assert, I strongly prefer cast for cases where we're just providing info to the type checker. If you're worried about overheads, I think it'd be relatively straightforward to statically remove all the casts in an import hook or so.

[alexfikl]

Found a way that avoids all the casting.

Nice! I thought of doing something like that too, but wasn't sure it would work :(

Btw, given the choice between cast and assert, I strongly prefer cast for cases where we're just providing info to the type checker.

Hm, I usually went by "if I need to cast, it probably means I didn't narrow the types enough for the type checker" and tried to avoid it. Yours is probably a better way to look at it when gradually adding types to things. :-?

[inducer]

Yours is probably a better way to look at it when gradually adding types to things. :-?

Not claiming it's a good way! And gradual typing mostly doesn't factor into it for the most part. Just trying to make sense of when to use what. Here's a different restatement:

• assert if the condition is a useful check to have at run time.
• cast for things the type checker needs to know but does not.

As a microbenchmark matter, the two have slightly different cost profiles, but I'm choosing to just regard both as "free", to avoid being pulled in spurious directions.

[inducer]

Thx!

[alexfikl]

Just trying to make sense of when to use what.

I'll happily defer to your newly gained extensive experience in typing 😁

As a microbenchmark matter, the two have slightly different cost profiles, but I'm choosing to just regard both as "free", to avoid being pulled in spurious directions.

I can definitely agree with that!