-
Notifications
You must be signed in to change notification settings - Fork 125
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
AVX support #590
Comments
Sounds good to me! We may need some changes in Rustler precompiled so we include this as a flag too. |
Hey @petrkozorezov , there is a PR adding this feature: #710 Would you mind to share your opinions? Do you think it covers what you proposed? I couldn't cover all the targets that run in x86_64 because some of them didn't compile with the flag (Apple one), or I couldn't easily detect if the target was running with the features needed enabled. But for the later, a compile env (or a env var) can be used to turn on the "legacy mode". WDYT? |
Looks great! But it's a little bit unclear (without going deeper into the code) how to use legacy mode build (and that it's even exists). Maybe add a few words about it to README? |
@petrkozorezov Thanks! I added a note to the README.md. I'm going to hold the PR until we decide to launch a new version, since this is going to be a breaking change due to the dependency on RustlerPrecompiled |
@philss we can merge if you want. tokenizers depends on ~> 0.6, so it should work with ~> 0.7 as well just fine. :) |
@josevalim done! Thanks! |
By default, polars python library builds with AVX/AVX2/... support and there is a separate version for legacy processors.
I checked explorer code and didn't found any mentions of AVX. So a tried to build explorer with RUSTFLAGS from polars and saw 4% performance increasing in my own code.
So maybe follow polars python library way: prebuild two versions of explorer (with (by default) and without AVX support) and switch between them with a flag (or smth like that)?
How I built explorer with AVX support:
UPD: Enabling AVX2 adds me another 1% of performance.
The text was updated successfully, but these errors were encountered: