Chorba: A novel CRC32 implementation (2024)(arxiv.org)
arxiv.org
Chorba: A novel CRC32 implementation (2024)
https://arxiv.org/abs/2412.16398
20 comments
This repo's readme gives a great overview of the previous-best approaches (as of ~6 years ago): https://github.com/komrad36/CRC
Warning: the README is 134KB... It's good but, like me, you might be here a while.
In Spain, "chorba" is very informal slang for "gal" [0]. Not vulgar, just very informal vernacular.
[0] https://dle.rae.es/chorbo
[0] https://dle.rae.es/chorbo
Chorba is also soup in Eastern European languages like Bulgarian and Romanian.
https://en.wikipedia.org/wiki/Chorba
https://en.wikipedia.org/wiki/Chorba
And also in Turkey. It (the word, if not the stew itself) arrived to the Balkans by way of the Ottomans. (And having just now clicked through to the link, it seems to have arrived to Turkey by way of the Persians).
Apparently that's kinda where the name comes from. It's named after a Serbian musician who was known as Bora Čorba, who played for a band called Riblja Čorba (fish stew).
All this talk of soup making me wonder if these are Arabic/ME derivatives.
Then again there are like 10 different ways to refer to soup in the various dialects.
Then again there are like 10 different ways to refer to soup in the various dialects.
The wikipedia article traces it to Persian, which formed it as a compound of words from different East Iranian languages. So you are on the money with Middle Eastern. From there it spread to the Balkan via Ottoman Turkish, and also from Persian to dialectal Arabic, which would explain the occurrences in Northern Africa, and maybe even Spain
and is a traditional Tunisian soup
/Edit: actually in all North Africa https://en.wikipedia.org/wiki/Chorba
/Edit: actually in all North Africa https://en.wikipedia.org/wiki/Chorba
Anyone can replicate the results? In any case, works like this give me moments of epiphany when I start to believe the humanity is not totally lost.
zlib-ng implements it so I assume it is practical at the very least
the zlib-ng guys are super welcoming and helpful, working with them was a very pleasant experience
Oh hey, the man himself!
I was looking at the zlib-ng crc32 implementation which is where I saw that it was recently updated to include your algorithm.
Good work, it's a surprisingly elegant solution when compared to the braiding approaches!
I was looking at the zlib-ng crc32 implementation which is where I saw that it was recently updated to include your algorithm.
Good work, it's a surprisingly elegant solution when compared to the braiding approaches!
thanks :) the braiding approach is super clever too, this was one of those weird moments where you find something and then have to triple check your results because how could i accidentally find something better than the algorithm that hasn't been touched in decades...
the part i really like is that it gives us small improvement on the pclmul too, as the non-accelerated algorithm doesn't really stand a chance against the accelerated opcode on newer hardware so it probably isn't going to see much use in practice. however... i think hardware solutions could possibly benefit (e.g. ethernet cards)
the part i really like is that it gives us small improvement on the pclmul too, as the non-accelerated algorithm doesn't really stand a chance against the accelerated opcode on newer hardware so it probably isn't going to see much use in practice. however... i think hardware solutions could possibly benefit (e.g. ethernet cards)
What are the units on the vertical axes for figures 1 and 2? I might have guessed seconds per TiB but the braiding line doesn't seem to match what's in figure 3.
I forget, these were the outputs of my test scripts and I wasn't exactly fastidious here, at least the axes all start from zero :)
IIRC they're off by maybe a factor of 10 or 100, the test scripts just generate a bunch of (seeded) random data and then execute the CRC of X bits Y times and that's where the number comes from, it's consistent across the different tests even if the units are wrong
IIRC they're off by maybe a factor of 10 or 100, the test scripts just generate a bunch of (seeded) random data and then execute the CRC of X bits Y times and that's where the number comes from, it's consistent across the different tests even if the units are wrong
News to me, but a guy named Sam Russell came up with a new software only CRC32 algorithm that is competitive with hardware accelerated implementations. It's a surprisingly elegant solution.
> This implementation is named after the Serbian singer Bora Đorđević (also known as Bora Čorba) who was born in 1952 and died in 2024. His birth year matches the number of the GZIP standard RFC 1952 that describes a common CRC32 implementation, and the original proof of concept for this method used the polynomial x21 +x15 + x14 + x11 + x10 + x7 + x3 which is x1952×8 mod G(x).
That is indeed dedication.