Faster Progress Report 2: Up to 600% speedups for popular crates, and our biggest release ever


#1

Hi everybody,

I’m proud to announce the fourth major release of faster, the SIMD library for humans. Many of the new features are detailed in faster’s second progress report, along with some of my attempts to integrate it into the greater Rust ecosystem.

As always, I’d lilke to extend my thanks to everybody in the Rust community for their support and assistance. Faster would not be nearly as fast and awesome as currently is without this community.


#2

It seems this library is growing well, better than expected.


#3

Also, faster than expected :wink:

Awesome work Mr Niederer, thank you for advancing this corner of rust!


#4

As I understand it, it supports data of plain natural (not simd) alignment by default. I’d love to look at how ndarray can use some of this, both internally and externally.

What’s your hunch, is there any utility to faster on data that is (possibly) not contiguous?


#5

I have a few functions which support noncontiguous data, but it’s all scalar at the moment. Check out PackedStripe in the docs. I should have vectorized noncontiguous loads/stores by 0.6.0, for *32s and *64s, and for almost-contiguous data (every other, every fourth, etc).