Skip to content

Conversation

@MilesCranmer
Copy link
Contributor

@MilesCranmer MilesCranmer commented Mar 20, 2024

I think some caching of metadata will help ReadOnlyArrays to start to get performance improvements over regular arrays.

Since a particular ReadOnlyArray cannot change shape, you can cache its shape information in the struct itself as an immutable value. Note that this is not stored as a type parameter – which is what StaticArrays.jl does – this change instead stores it as a value. But, since this value is immutable, LLVM should be able to do things like infer that two loops, one after the other, have the same number of iterations, and do some optimizations with that information.

While this does not yet take advantage of the fact that the contents are unchanging (see JuliaLang/julia#44381 for that), it will hopefully get some perf improvements from cached shape info.

@MilesCranmer
Copy link
Contributor Author

MilesCranmer commented Mar 20, 2024

The only tricky thing is that if the user modifies the underlying array (directly), these assumptions will get violated...

@bkamins
Copy link
Collaborator

bkamins commented Mar 27, 2024

The original idea of ReadOnlyArrays.jl was to protect from unwanted mutation. As you comment we currently do not guarantee that the underlying array will not change.

@MilesCranmer
Copy link
Contributor Author

Yeah good point. I guess this is better left for a core array operation (JuliaLang/julia#44381 — though that has stalled).

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants