Improve performance of `rem_scalar/div_scalar` for integer types (4x-10x) about arrow2 HOT 7 CLOSED

Fuse processing number% 3 is slow. The main reason is that type conversion is made every time. I try to remove type conversion. The performance can be 2G -- > 5G / s

from arrow2.

@lideen999

#252 shows the main reason of cast.

I did some perfs in datafuse, seems the rem is the hot path.

from arrow2.

@jorgecarleitao

By using strength_reduce, it shows great improvement in datafuse.

from arrow2.

This is a valid request: valid use-case, documented benefits, crate with implementation available. 👍

Would you like to work on it, or would you like me to take it?

Also, I went through the crate and is unsafe free, so even easier sell.

from arrow2.

The same optimization can be used for division as well right?

from arrow2.

The same optimization can be used for division as well right?

Yes! Of course.

Would you like to work on it, or would you like me to take it?

Sorry, currently I may have no time to work on it, because there are some urgent issues I have to do in datafuse.

Now I just provide this idea to make arrow2 work better.
And strength_reduce is not generically implemented, so it may need some type match case to dispatch the codes.

from arrow2.

Done in #275

from arrow2.