Bernat Font

~Ah but this is not a problem of this PR, but a problem of WaterLily-Benchmarks, right? If you open an issue there, we can iterate on it.~ You mean these...

As a result of our conversation in https://github.com/WaterLily-jl/WaterLily.jl/pull/198, I thought it was about time to put this in use... So I have cleaned up a bit the Preferences.jl routines with...

To do: - [x] Automate CI with single thread and Preferences - [x] Check again performance of not specifying workgroup size VS `64` (master) VS `(64,64)` - [ ] Get...

I have experimenting again with the workgroup size, and the best results are almost always with constant `64`. Also, something that I do not understand, is that the single-thread benchmarks...

New CI working nicely with LocalPreferences!

@vchuravy any ideas on how to bypass the warning of using KA in single thread (instead of SIMD) during precompilation? Otherwise, since we use KA as default backend and precompilation...

The `@loop` macro now implements automatic specialization for the function wrapping the kernels, so that all arguments are included such as ```julia function kern(a::A,b::B,c::C,...) where {A,B,C,...} ... end ``` Below...

Consistent 1-2% speedup on all backends, and we will (hopefully) be able to specialize kernels passing a function. So this ticks all the boxes :) Benchmarks ``` Benchmark environment: tgv...

We should address the warning issue, and this PR is good to go!

Yes! Removing the dynamic dispatch based on number of threads, and instead just compiling the SIMD kernel based on LocalPreferences, brought down allocations significantly for single thread. Then the general...