Shane F. Carr

> Now the forward pass can be expressed in the form of matrix multiplication. and elementwise dot product (which might be vectorized). I just want to clarify . `dot_3d` is...

@pdogr If you can resolve the remaining open questions on this PR in the next ~day, it would be nice to land it so that we can ship it in...

Current benches on M2 MacBook Air: ``` # Main Line Break/UTF8/Th/lstm time: [231.19 µs 231.31 µs 231.43 µs] # sffc/vectorize-convolve Line Break/UTF8/Th/lstm time: [215.82 µs 216.38 µs 217.30 µs] #...

Please note that the memory layout of `sffc/vectorize-convolve` and `pdogr/2d-mat` is the same, so careful coding should be able to bring the performance together without touching the data model: https://raw.githubusercontent.com/unicode-org/icu4x/532a71bc79a0c3d4bab0761a1ee3499256528bf0/provider/repodata/data/json/segmenter/lstm/wl_auto%401/th.json...

I'd be happy to see whether @pdogr can get back to the 203.37 µs. It should be achievable without touching the data model.

Eager to see how much performance you can get out of this. I will reiterate my position though that I don't currently see a logical reason why a 4*HxD should...

Why can't we use `blis` with the upgraded 3-D type? You can just drop the third dimension when you call into the library or perform any other operation. What I'm...

> * [ ] There are 6 different formatters - are they all needed? They support different field sets, so you would need all the ones that correspond to field...

> > Because they load different data. The match statement you wrote looks correct. > > But that can lead to errors, when users do things like (pseudocode) `ZonedDateTimeFormatter.specificShort('America/Chicago')`. Is...

This PR broke `bench-binsize` CI ``` error: macro expansion ends with an incomplete expression: expected expression --> components/icu/examples/tui.rs:82:36 | 21 | ($($arg:tt)*) => {}; | -- in this macro arm...