Making CRC calculations in Mojo 18x faster than Python and 3x slower than Python | Hacker News Reader

Making CRC calculations in Mojo 18x faster than Python and 3x slower than Python

fnands.com

58 pointsfnands17 comments

For those that thought the process of speeding up CRC was interesting, I strongly recommend reading [1]. It describes a step by step process on how a naive CRC implementation might be improved, until finally arriving at an implementation in assembly with a staggering throughput of 62 processed bits (almost 8 bytes) per CPU cycle. Yes, you read that right.

[1]: https://github.com/komrad36/CRC

Yup, it's a fantastic read. I based most of my post off it (I clearly mention so) and it's worth it to read at least the first part of it before reading my post.

For what it's worth, it appears the paper "Everything we know about CRC but afraid to forget" was originally published as part of the release of crcutil on Google Code[1]. This is a hg repository with one commit that includes the paper, the source of the paper, and an implementation.

[1]: https://code.google.com/archive/p/crcutil/

Thanks! I'll add it to the post.

FYI, I forked and improved [1] a Rust implementation that supports both table- and SIMD-accelerated CRC-64/NVME [2] calculations. The SIMD-accelerated (x86/x86_64 and aarch64) version delivers 10X over the table (16-bytes at a time) implementation.

The original implementation [3] did the same thing but for CRC-64/XZ [4].

[1]: https://github.com/awesomized/crc64fast-nvme

[2]: https://reveng.sourceforge.io/crc-catalogue/all.htm#crc.cat....

[3]: https://github.com/tikv/crc64fast

[4]: https://reveng.sourceforge.io/crc-catalogue/all.htm#crc.cat....

How does it compare to the built-in CRC32 instruction? [1]

[1] https://www.intel.com/content/www/us/en/docs/ipp/developer-g...

This is computing CRC-64, not CRC-32, so there's not really a comparison. But perhaps most importantly, ours works with a variety of polynomials (there are a lot! [1])... we're just using the NVME one, but it's trivially adaptable to most (all?) of them. (The Intel instruction you link to only works with two - CRC32 and CRC32C)

Finally, it's based on Intel's paper [2], so they also believe it's extremely fast. :)

[1]: https://reveng.sourceforge.io/crc-catalogue/all.htm

[2]: https://web.archive.org/web/20131224125630/https://www.intel...

https://github.com/py2many/py2many/pull/653

Transpiling the python version in the blog to mojo gives me a 4x speedup.

Had to hand edit a couple of things:

    - List to bytearray conversion is not working yet
    - Can't iterate over SIMD. So a "for x in list" loop has to be rewritten as a range based loop.

With a bit more work, the manual edits won't be necessary.

Generated code: https://gist.github.com/adsharma/22ec18664ce4a59750b78fddc58...

Edited code: https://gist.github.com/adsharma/72516d8333a25fd8e350472568d...

Oh! Cool project!

You mean 4x speedup over Python?

Yes. Details in the pull request linked above

How is mojo doing ? Has it made its way as a niche language in some places ?

It's coming along. I don't think anyone is using it for anything serious yet, but it is starting to feel like a real language.

My guess is that it will start being used as a library language (i.e. have libraries written in Mojo being called from Python) before it really gets going as its own thing.

and is it still closed source?

Yup, language is still closed, stdlib is open.

Great post @fnands!

Thanks Chris!

It has been about a year since I started playing around with Mojo and it's impressive how far it (and the community) has come!

Reply on news.ycombinator.com