Rework

[9il]

Changelog:

1. Reworked sweep_slice to access memory only once like the Python code. I didn't make a benchmark, but I expect it will be a few times faster now.
2. Added fastmath to the kernel-like API and pragma(inline, false) for some API. The last one is required to make the first one work well.
3. Changed / with * where possible (it is fine for this kind of numeric code).
4. Used more mir.* alternatives if possible.
5. Simplified API for some functions
6. ndslice-like rework for restriction
7. ndslice-like rework for prolongation
8. @nogc cycle

[9il]

Hi @typohnebild
Please let me know how the numbers look like when you have a chance.
I don't have a Linux machine to run the scripts :/

[typohnebild]

Hi @9il,
thank you very much for your rework!
I have a done a small comparison between your branch and ours. Your rework gives a enormous speed up for all variations.
Especially the slice version only takes have the time

see
see

raw
raw

[9il]

Thanks!
Hope I will add sweep_ndslice next week.
sweep_ndslicewill have a mir.ndslice idiomatic implementation like the new projection, also weighted projection can be additionally simplified.

And maybe a multithreaded version will be added too.

Would you mind to update the README and charts when all changes ready?

[sm1th3ry]

This is a great job you are doing! Thanks a lot!
We are learning a lot about D by analysing your code.

Of course. We will rerun the benchmarks when everything is done and update the README. Just let us know when you are finished.

[9il]

done

PoissonCycle!(double, Dim, SweepType.ndslice) can be built with up to 8-dimensional space, higher dimensions may fail to compile.

[9il]

The default sweep algorithm changes to ndslice, which may not be the fastest version. It is smaller, either to write and supports higher dimensions.

The field version doesn't use Mir except for API and looks more like C/C++ code. It isn't bad, but I am afraid that it may scare away potential users if we make it athe default sweep algorithm.

[typohnebild]

Thanks a lot!

We are going to merge this and then start a new run and afterwards we will update the readme!