Add a couple of precompiles

[timholy]

No description provided.

[yurivish]

Hi Tim! I'm wondering – what does executing these precompile statements get you beyond what you'd get from calling the functions for the first time, which (IIUC) would compile them at the time of first use?

[timholy]

Lower latency on first use. I didn't measure here, but for reference: JuliaImages/ImageFiltering.jl#201.

I'm not sure the ones here "work," though; there are a few cases that don't, for reasons the community has yet to fully understand. (E.g., JuliaLang/julia#38983.) But eventually they will start working, so it's probably not bad to have them.

[yurivish]

Thanks for the links! Just to make sure I understand, the tradeoff here is to incur the latency cost every time the package is loaded, regardless of whether any of these functions is used during the session, rather than incurring it at the first point of use? I haven’t been following the recent compiler improvements but I know there are varying levels of precompilation.

I can see why this might be preferable since one can go and make coffee or someone while the “using” statement is being executed...

[timholy]

the tradeoff here is to incur the latency cost every time the package is loaded,

No, the cost of precompile is paid just when it's precompiled. I.e., when the package is being built. (Precompiling SignedDistanceFields...) There's a small extra penalty to load the type-inferred MethodInstances but that's much smaller than needing to run inference every time you use the methods.

In other words, this is Julia's closest equivalent to make in the C/C++ world, where once you've compiled the application it's fast to launch.

[yurivish]

Ah, got it. Thanks!

[timholy]

Your question was a good one, though; I remember thinking the same thing. It was also timely, as it prompted me to explain this specific point in JuliaLang/www.julialang.org#1111. (In the preview, scroll down to "It might be natural to wonder...")