Support for Fat Teddy in aarch64?

[arunj18]

Hello!

I want to start off by saying that we are avid users of this package and love the performance (along with the optimizations!) of this package!

Now, providing some background, we run a system that matches multiple patterns on textual data we process (cannot share the data as it is sensitive). We use aho-corasick in our Rust code to perform some of the matching by using the find_overlapping_iter method on a string while searching for multiple patterns.

Recently, we have decided to move some of our workloads from X86_64 based compute to aarch64 based compute where we've seen degradation in performance when using aho-corasick.

Our main theory at the moment after looking through the repository is the lack of a "Fat" version of the Teddy prefilter for the aarch64 platform. Are there any plans to add an implementation for it?

Happy to provide some metrics on the changes we've observed but unable to share the code/our usage of it/data at the moment.

[BurntSushi]

Were your programs hitting Teddy on x86-64? Have you confirmed that by looking at a profile?

In any case, no, there aren't any plans because it doesn't really make sense. x86-64 has a fat Teddy because you can use wider vectors there (32-bit). But aarch64 is limited to 16-bit vectors, which makes fat Teddy infeasible as far as I know.

[arunj18]

Let me see if I can confirm that Fat Teddy is being used in our code in the X86_64 (profiling requires some additional work).

Does the aarch64 limitation of 16-bit vectors still hold with addition of Scalable Vector Extensions (https://developer.arm.com/documentation/102476/latest/) in newer revisions? Apologies if that doesn't make sense, I have some knowledge of ARM instructions but not much knowledge about SIMD - ARM instructions.

[BurntSushi]

AFAIK, SVE isn't widely supported. And Rust doesn't have an answer for it even if it was.

I'm not an aarch64 expert. You may need to do your own leg work here.